Development and Validation of a Deep Learning System for Diabetic Retinopathy and Related Eye Diseases Using Retinal Images From Multiethnic Populations With Diabetes

Building a new regional home for implementation science: Annual Midwest Clinical & Translational Research Meetings

The vision of the Central Society for Clinical and Translational Research (CSCTR) is to "promote a vibrant, supportive community of multidisciplinary, clinical, and translational medical research to benefit humanity." Together with the Midwestern Section of the American Federation for Medical Research, CSCTR hosts an Annual Midwest Clinical & Translational Research Meeting, a regional multispecialty meeting that provides the opportunity for trainees and early-stage investigators to present their research to leaders in their fields. There is an increasing national and global interest in implementation science (IS), the systematic study of activities (or strategies) to facilitate the successful uptake of evidence-based health interventions in clinical and community settings. Given the growing importance of this field and its relevance to the goals of the CSCTR, in 2022, the Midwest Clinical & Translational Research Meeting incorporated new initiatives and sessions in IS. In this report, we describe the role of IS in the translational research spectrum, provide a summary of sessions from the 2022 Midwest Clinical & Translational Research Meeting, and highlight initiatives to complement national efforts to build capacity for IS through the annual meetings.

Artificial intelligence for diabetic retinopathy in low-income and middle-income countries: a scoping review

Diabetic retinopathy (DR) is a leading cause of blindness globally. There is growing evidence to support the use of artificial intelligence (AI) in diabetic eye care, particularly for screening populations at risk of sight loss from DR in low-income and middle-income countries (LMICs) where resources are most stretched. However, implementation into clinical practice remains limited. We conducted a scoping review to identify what AI tools have been used for DR in LMICs and to report their performance and relevant characteristics. 81 articles were included. The reported sensitivities and specificities were generally high providing evidence to support use in clinical practice. However, the majority of studies focused on sensitivity and specificity only and there was limited information on cost, regulatory approvals and whether the use of AI improved health outcomes. Further research that goes beyond reporting sensitivities and specificities is needed prior to wider implementation.

Towards population-independent, multi-disease detection in fundus photographs

Independent validation studies of automatic diabetic retinopathy screening systems have recently shown a drop of screening performance on external data. Beyond diabetic retinopathy, this study investigates the generalizability of deep learning (DL) algorithms for screening various ocular anomalies in fundus photographs, across heterogeneous populations and imaging protocols. The following datasets are considered: OPHDIAT (France, diabetic population), OphtaMaine (France, general population), RIADD (India, general population) and ODIR (China, general population). Two multi-disease DL algorithms were developed: a Single-Dataset (SD) network, trained on the largest dataset (OPHDIAT), and a Multiple-Dataset (MD) network, trained on multiple datasets simultaneously. To assess their generalizability, both algorithms were evaluated whenever training and test data originate from overlapping datasets or from disjoint datasets. The SD network achieved a mean per-disease area under the receiver operating characteristic curve (mAUC) of 0.9571 on OPHDIAT. However, it generalized poorly to the other three datasets (mAUC < 0.9). When all four datasets were involved in training, the MD network significantly outperformed the SD network (p = 0.0058), indicating improved generality. However, in leave-one-dataset-out experiments, performance of the MD network was significantly lower on populations unseen during training than on populations involved in training (p < 0.0001), indicating imperfect generalizability.

Performance of a Support Vector Machine Learning Tool for Diagnosing Diabetic Retinopathy in Clinical Practice

PURPOSE

To examine the real-world performance of a support vector machine learning software (RetinaLyze) in order to identify the possible presence of diabetic retinopathy (DR) in patients with diabetes via software implementation in clinical practice.

METHODS

1001 eyes from 1001 patients-one eye per patient-participating in the Danish National Screening Programme were included. Three independent ophthalmologists graded all eyes according to the International Clinical Diabetic Retinopathy Disease Severity Scale with the exact level of disease being determined by majority decision. The software detected DR and no DR and was compared to the ophthalmologists' gradings.

RESULTS

At a clinical chosen threshold, the software showed a sensitivity, specificity, positive predictive value and negative predictive value of 84.9% (95% CI: 81.8-87.9), 89.9% (95% CI: 86.8-92.7), 92.1% (95% CI: 89.7-94.4), and 81.0% (95% CI: 77.2-84.7), respectively, when compared to human grading. The results from the routine screening were 87.0% (95% CI: 84.2-89.7), 85.3% (95% CI: 81.8-88.6), 89.2% (95% CI: 86.3-91.7), and 82.5% (95% CI: 78.5-86.0), respectively. AUC was 93.4%. The reference graders Conger's Exact Kappa was 0.827.

CONCLUSION

The software performed similarly to routine grading with overlapping confidence intervals, indicating comparable performance between the two groups. The intergrader agreement was satisfactory. However, evaluating the updated software alongside updated clinical procedures is crucial. It is therefore recommended that further clinical testing before implementation of the software as a decision support tool is conducted.

Artificial intelligence for fracture diagnosis in orthopedic X-rays: current developments and future potential

The use of artificial intelligence (AI) in the interpretation of orthopedic X-rays has shown great potential to improve the accuracy and efficiency of fracture diagnosis. AI algorithms rely on large datasets of annotated images to learn how to accurately classify and diagnose abnormalities. One way to improve AI interpretation of X-rays is to increase the size and quality of the datasets used for training, and to incorporate more advanced machine learning techniques, such as deep reinforcement learning, into the algorithms. Another approach is to integrate AI algorithms with other imaging modalities, such as computed tomography (CT) scans, and magnetic resonance imaging (MRI), to provide a more comprehensive and accurate diagnosis. Recent studies have shown that AI algorithms can accurately detect and classify fractures of the wrist and long bones on X-ray images, demonstrating the potential of AI to improve the accuracy and efficiency of fracture diagnosis. These findings suggest that AI has the potential to significantly improve patient outcomes in the field of orthopedics.

Expert and deep learning model identification of iEEG seizures and seizure onset times

Hundreds of 90-s iEEG records are typically captured from each NeuroPace RNS System patient between clinic visits. While these records provide invaluable information about the patient's electrographic seizure and interictal activity patterns, manually classifying them into electrographic seizure/non-seizure activity, and manually identifying the seizure onset channels and times is an extremely time-consuming process. A convolutional neural network based Electrographic Seizure Classifier (ESC) model was developed in an earlier study. In this study, the classification model is tested against iEEG annotations provided by three expert reviewers board certified in epilepsy. The three experts individually annotated 3,874 iEEG channels from 36, 29, and 35 patients with leads in the mesiotemporal (MTL), neocortical (NEO), and MTL + NEO regions, respectively. The ESC model's seizure/non-seizure classification scores agreed with the three reviewers at 88.7%, 89.6%, and 84.3% which was similar to how reviewers agreed with each other (92.9%-86.4%). On iEEG channels with all 3 experts in agreement (83.2%), the ESC model had an agreement score of 93.2%. Additionally, the ESC model's certainty scores reflected combined reviewer certainty scores. When 0, 1, 2 and 3 (out of 3) reviewers annotated iEEG channels as electrographic seizures, the ESC model's seizure certainty scores were in the range: [0.12-0.19], [0.32-0.42], [0.61-0.70], and [0.92-0.95] respectively. The ESC model was used as a starting-point model for training a second Seizure Onset Detection (SOD) model. For this task, seizure onset times were manually annotated on a relatively small number of iEEG channels (4,859 from 50 patients). Experiments showed that fine-tuning the ESC models with augmented data (30,768 iEEG channels) resulted in a better validation performance (on 20% of the manually annotated data) compared to training with only the original data (3.1s vs 4.4s median absolute error). Similarly, using the ESC model weights as the starting point for fine-tuning instead of other model weight initialization methods provided significant advantage in SOD model validation performance (3.1s vs 4.7s and 3.5s median absolute error). Finally, on iEEG channels where three expert annotations of seizure onset times were within 1.5 s, the SOD model's seizure onset time prediction was within 1.7 s of expert annotation.

The Definition of Glaucomatous Optic Neuropathy in Artificial Intelligence Research and Clinical Applications

OBJECTIVE

Although artificial intelligence (AI) models may offer innovative and powerful ways to use the wealth of data generated by diagnostic tools, there are important challenges related to their development and validation. Most notable is the lack of a perfect reference standard for glaucomatous optic neuropathy (GON). Because AI models are trained to predict presence of glaucoma or its progression, they generally rely on a reference standard that is used to train the model and assess its validity. If an improper reference standard is used, the model may be trained to detect or predict something that has little or no clinical value. This article summarizes the issues and discussions related to the definition of GON in AI applications as presented by the Glaucoma Workgroup from the Collaborative Community for Ophthalmic Imaging (CCOI) US Food and Drug Administration Virtual Workshop, on September 3 and 4, 2020, and on January 28, 2022.

DESIGN

Review and conference proceedings.

SUBJECTS

No human or animal subjects or data therefrom were used in the production of this article.

METHODS

A summary of the Workshop was produced with input and approval from all participants.

MAIN OUTCOME MEASURES

Consensus position of the CCOI Workgroup on the challenges in defining GON and possible solutions.

RESULTS

The Workshop reviewed existing challenges that arise from the use of subjective definitions of GON and highlighted the need for a more objective approach to characterize GON that could facilitate replication and comparability of AI studies and allow for better clinical validation of proposed AI tools. Different tests and combination of parameters for defining a reference standard for GON have been proposed. Different reference standards may need to be considered depending on the scenario in which the AI models are going to be applied, such as community-based or opportunistic screening versus detection or monitoring of glaucoma in tertiary care.

CONCLUSIONS

The development and validation of new AI-based diagnostic tests should be based on rigorous methodology with clear determination of how the reference standards for glaucomatous damage are constructed and the settings where the tests are going to be applied.

FINANCIAL DISCLOSURE(S)

Proprietary or commercial disclosure may be found after the references.

The Impact of Artificial Intelligence on the Preference of Radiology as a Future Specialty Among Medical Students at Jazan University, Saudi Arabia: A Cross-Sectional Study

Background The use of artificial intelligence (AI) in healthcare continues to spark interest and has been the subject of extensive discussion in recent years as well as its potential effects on future medical specialties, including radiology. In this study, we aimed to study the impact of AI on the preference of medical students at Jazan University in choosing radiology as a future specialty. Methodology An observational cross-sectional study was conducted using a pre-tested self-administered online questionnaire among medical students at Jazan University. Data were cleaned, coded, entered, and analyzed using SPSS (SPSS Inc., USA) version 25. Statistical significance was defined as a P-value of less than 0.05. We examined the respondents' preference for radiology rankings with the presence and absence of AI. Radiology's ranking as a preferred specialty with or without AI integration was statistically analyzed for associations with baseline characteristics, personal opinions, and previous exposures among those who had radiology as one of their top three options. Results Approximately 27.4% of males and 28.3% of females ranked radiology among their top three preferred choices. Almost 65.2% were exposed to radiology topics through pre-clinical lectures. The main sources of information about AI for the studied group were medical students (41%) and the Internet (27.5%). The preference of students for radiology was significantly affected when it is assessed by AI (P < 0.05). Around (16.1%) of those who chose radiology as one of their top three choices strongly agree that AI will decrease the job opportunities for radiologists. Logistic regression analysis showed that being a female is significantly associated with an increased chance to replace radiology with other specialty when it is integrated with AI (Crude odds ratio (COR) = 1.91). Conclusion Our results demonstrated that the students' choices were significantly affected by the presence of AI. Thereover, to raise medical students' knowledge and awareness of the potential positive effects of AI, it is necessary to organize an educational campaign, webinars, and conferences.

Artificial intelligence in ophthalmology: The path to the real-world clinic

Artificial intelligence (AI) has great potential to transform healthcare by enhancing the workflow and productivity of clinicians, enabling existing staff to serve more patients, improving patient outcomes, and reducing health disparities. In the field of ophthalmology, AI systems have shown performance comparable with or even better than experienced ophthalmologists in tasks such as diabetic retinopathy detection and grading. However, despite these quite good results, very few AI systems have been deployed in real-world clinical settings, challenging the true value of these systems. This review provides an overview of the current main AI applications in ophthalmology, describes the challenges that need to be overcome prior to clinical implementation of the AI systems, and discusses the strategies that may pave the way to the clinical translation of these systems.

Synthetic artificial intelligence using generative adversarial network for retinal imaging in detection of age-related macular degeneration

Introduction

Age-related macular degeneration (AMD) is one of the leading causes of vision impairment globally and early detection is crucial to prevent vision loss. However, the screening of AMD is resource dependent and demands experienced healthcare providers. Recently, deep learning (DL) systems have shown the potential for effective detection of various eye diseases from retinal fundus images, but the development of such robust systems requires a large amount of datasets, which could be limited by prevalence of the disease and privacy of patient. As in the case of AMD, the advanced phenotype is often scarce for conducting DL analysis, which may be tackled via generating synthetic images using Generative Adversarial Networks (GANs). This study aims to develop GAN-synthesized fundus photos with AMD lesions, and to assess the realness of these images with an objective scale.

Methods

To build our GAN models, a total of 125,012 fundus photos were used from a real-world non-AMD phenotypical dataset. StyleGAN2 and human-in-the-loop (HITL) method were then applied to synthesize fundus images with AMD features. To objectively assess the quality of the synthesized images, we proposed a novel realness scale based on the frequency of the broken vessels observed in the fundus photos. Four residents conducted two rounds of gradings on 300 images to distinguish real from synthetic images, based on their subjective impression and the objective scale respectively.

Results and discussion

The introduction of HITL training increased the percentage of synthetic images with AMD lesions, despite the limited number of AMD images in the initial training dataset. Qualitatively, the synthesized images have been proven to be robust in that our residents had limited ability to distinguish real from synthetic ones, as evidenced by an overall accuracy of 0.66 (95% CI: 0.61-0.66) and Cohen's kappa of 0.320. For the non-referable AMD classes (no or early AMD), the accuracy was only 0.51. With the objective scale, the overall accuracy improved to 0.72. In conclusion, GAN models built with HITL training are capable of producing realistic-looking fundus images that could fool human experts, while our objective realness scale based on broken vessels can help identifying the synthetic fundus photos.

Relationships between quantitative retinal microvascular characteristics and cognitive function based on automated artificial intelligence measurements

Introduction: The purpose of this study is to assess the relationship between retinal vascular characteristics and cognitive function using artificial intelligence techniques to obtain fully automated quantitative measurements of retinal vascular morphological parameters. Methods: A deep learning-based semantic segmentation network ResNet101-UNet was used to construct a vascular segmentation model for fully automated quantitative measurement of retinal vascular parameters on fundus photographs. Retinal photographs centered on the optic disc of 3107 participants (aged 50-93 years) from the Beijing Eye Study 2011, a population-based cross-sectional study, were analyzed. The main parameters included the retinal vascular branching angle, vascular fractal dimension, vascular diameter, vascular tortuosity, and vascular density. Cognitive function was assessed using the Mini-Mental State Examination (MMSE). Results: The results showed that the mean MMSE score was 26.34 ± 3.64 (median: 27; range: 2-30). Among the participants, 414 (13.3%) were classified as having cognitive impairment (MMSE score < 24), 296 (9.5%) were classified as mild cognitive impairment (MMSE: 19-23), 98 (3.2%) were classified as moderate cognitive impairment (MMSE: 10-18), and 20 (0.6%) were classified as severe cognitive impairment (MMSE < 10). Compared with the normal cognitive function group, the retinal venular average diameter was significantly larger (p = 0.013), and the retinal vascular fractal dimension and vascular density were significantly smaller (both p < 0.001) in the mild cognitive impairment group. The retinal arteriole-to-venular ratio (p = 0.003) and vascular fractal dimension (p = 0.033) were significantly decreased in the severe cognitive impairment group compared to the mild cognitive impairment group. In the multivariate analysis, better cognition (i.e., higher MMSE score) was significantly associated with higher retinal vascular fractal dimension (b = 0.134, p = 0.043) and higher retinal vascular density (b = 0.152, p = 0.023) after adjustment for age, best corrected visual acuity (BCVA) (logMAR) and education level. Discussion: In conclusion, our findings derived from an artificial intelligence-based fully automated retinal vascular parameter measurement method showed that several retinal vascular morphological parameters were correlated with cognitive impairment. The decrease in retinal vascular fractal dimension and decreased vascular density may serve as candidate biomarkers for early identification of cognitive impairment. The observed reduction in the retinal arteriole-to-venular ratio occurs in the late stages of cognitive impairment.

Management of keratoconus: an updated review

Keratoconus is the most common corneal ectatic disorder. It is characterized by progressive corneal thinning with resultant irregular astigmatism and myopia. Its prevalence has been estimated at 1:375 to 1:2,000 people globally, with a considerably higher rate in the younger populations. Over the past two decades, there was a paradigm shift in the management of keratoconus. The treatment has expanded significantly from conservative management (e.g., spectacles and contact lenses wear) and penetrating keratoplasty to many other therapeutic and refractive modalities, including corneal cross-linking (with various protocols/techniques), combined CXL-keratorefractive surgeries, intracorneal ring segments, anterior lamellar keratoplasty, and more recently, Bowman's layer transplantation, stromal keratophakia, and stromal regeneration. Several recent large genome-wide association studies (GWAS) have identified important genetic mutations relevant to keratoconus, facilitating the development of potential gene therapy targeting keratoconus and halting the disease progression. In addition, attempts have been made to leverage the power of artificial intelligence-assisted algorithms in enabling earlier detection and progression prediction in keratoconus. In this review, we provide a comprehensive overview of the current and emerging treatment of keratoconus and propose a treatment algorithm for systematically guiding the management of this common clinical entity.

A comparison of performance between a deep learning model with residents for localization and classification of intracranial hemorrhage

Intracranial hemorrhage (ICH) from traumatic brain injury (TBI) requires prompt radiological investigation and recognition by physicians. Computed tomography (CT) scanning is the investigation of choice for TBI and has become increasingly utilized under the shortage of trained radiology personnel. It is anticipated that deep learning models will be a promising solution for the generation of timely and accurate radiology reports. Our study examines the diagnostic performance of a deep learning model and compares the performance of that with detection, localization and classification of traumatic ICHs involving radiology, emergency medicine, and neurosurgery residents. Our results demonstrate that the high level of accuracy achieved by the deep learning model, (0.89), outperforms the residents with regard to sensitivity (0.82) but still lacks behind in specificity (0.90). Overall, our study suggests that the deep learning model may serve as a potential screening tool aiding the interpretation of head CT scans among traumatic brain injury patients.

Performance of artificial intelligence in diabetic retinopathy screening: a systematic review and meta-analysis of prospective studies

Aims

To systematically evaluate the diagnostic value of an artificial intelligence (AI) algorithm model for various types of diabetic retinopathy (DR) in prospective studies over the previous five years, and to explore the factors affecting its diagnostic effectiveness.

Materials and methods

A search was conducted in Cochrane Library, Embase, Web of Science, PubMed, and IEEE databases to collect prospective studies on AI models for the diagnosis of DR from January 2017 to December 2022. We used QUADAS-2 to evaluate the risk of bias in the included studies. Meta-analysis was performed using MetaDiSc and STATA 14.0 software to calculate the combined sensitivity, specificity, positive likelihood ratio, and negative likelihood ratio of various types of DR. Diagnostic odds ratios, summary receiver operating characteristic (SROC) plots, coupled forest plots, and subgroup analysis were performed according to the DR categories, patient source, region of study, and quality of literature, image, and algorithm.

Results

Finally, 21 studies were included. Meta-analysis showed that the pooled sensitivity, specificity, pooled positive likelihood ratio, pooled negative likelihood ratio, area under the curve, Cochrane Q index, and pooled diagnostic odds ratio of AI model for the diagnosis of DR were 0.880 (0.875-0.884), 0.912 (0.99-0.913), 13.021 (10.738-15.789), 0.083 (0.061-0.112), 0.9798, 0.9388, and 206.80 (124.82-342.63), respectively. The DR categories, patient source, region of study, sample size, quality of literature, image, and algorithm may affect the diagnostic efficiency of AI for DR.

Conclusion

AI model has a clear diagnostic value for DR, but it is influenced by many factors that deserve further study.

Systematic review registration

https://www.crd.york.ac.uk/prospero/, identifier CRD42023389687.

Torsional nystagmus recognition based on deep learning for vertigo diagnosis

Introduction

Detection of torsional nystagmus can help identify the canal of origin in benign paroxysmal positional vertigo (BPPV). Most currently available pupil trackers do not detect torsional nystagmus. In view of this, a new deep learning network model was designed for the determination of torsional nystagmus.

Methods

The data set comes from the Eye, Ear, Nose and Throat (Eye&ENT) Hospital of Fudan University. In the process of data acquisition, the infrared videos were obtained from eye movement recorder. The dataset contains 24521 nystagmus videos. All torsion nystagmus videos were annotated by the ophthalmologist of the hospital. 80% of the data set was used to train the model, and 20% was used to test.

Results

Experiments indicate that the designed method can effectively identify torsional nystagmus. Compared with other methods, it has high recognition accuracy. It can realize the automatic recognition of torsional nystagmus and provides support for the posterior and anterior canal BPPV diagnosis.

Discussion

Our present work complements existing methods of 2D nystagmus analysis and could improve the diagnostic capabilities of VNG in multiple vestibular disorders. To automatically pick BPV requires detection of nystagmus in all 3 planes and identification of a paroxysm. This is the next research work to be carried out.

Unmet needs and future perspectives in hydroxychloroquine retinopathy

Retinopathy is a well-recognized toxic effect of hydroxychloroquine treatment. As hydroxychloroquine retinopathy is potentially a vision-threatening condition, early detection is imperative to minimize vision loss due to drug toxicity. However, early detection of hydroxychloroquine retinopathy is still challenging even with modern retinal imaging techniques. No treatment has been established for this condition, except for drug cessation to minimize further damage. In this perspective article, we aimed to summarize the knowledge gaps and unmet needs in current clinical practice and research in hydroxychloroquine retinopathy. The information presented in this article may help guide the future directions of screening practices and research in hydroxychloroquine retinopathy.

Enhanced Deep Learning Model for Classification of Retinal Optical Coherence Tomography Images

Retinal optical coherence tomography (OCT) imaging is a valuable tool for assessing the condition of the back part of the eye. The condition has a great effect on the specificity of diagnosis, the monitoring of many physiological and pathological procedures, and the response and evaluation of therapeutic effectiveness in various fields of clinical practices, including primary eye diseases and systemic diseases such as diabetes. Therefore, precise diagnosis, classification, and automated image analysis models are crucial. In this paper, we propose an enhanced optical coherence tomography (EOCT) model to classify retinal OCT based on modified ResNet (50) and random forest algorithms, which are used in the proposed study's training strategy to enhance performance. The Adam optimizer is applied during the training process to increase the efficiency of the ResNet (50) model compared with the common pre-trained models, such as spatial separable convolutions and visual geometry group (VGG) (16). The experimentation results show that the sensitivity, specificity, precision, negative predictive value, false discovery rate, false negative rate accuracy, and Matthew's correlation coefficient are 0.9836, 0.9615, 0.9740, 0.9756, 0.0385, 0.0260, 0.0164, 0.9747, 0.9788, and 0.9474, respectively.

Deep learning for automated detection of neovascular leakage on ultra-widefield fluorescein angiography in diabetic retinopathy

Diabetic retinopathy is a leading cause of blindness in working-age adults worldwide. Neovascular leakage on fluorescein angiography indicates progression to the proliferative stage of diabetic retinopathy, which is an important distinction that requires timely ophthalmic intervention with laser or intravitreal injection treatment to reduce the risk of severe, permanent vision loss. In this study, we developed a deep learning algorithm to detect neovascular leakage on ultra-widefield fluorescein angiography images obtained from patients with diabetic retinopathy. The algorithm, an ensemble of three convolutional neural networks, was able to accurately classify neovascular leakage and distinguish this disease marker from other angiographic disease features. With additional real-world validation and testing, our algorithm could facilitate identification of neovascular leakage in the clinical setting, allowing timely intervention to reduce the burden of blinding diabetic eye disease.

A knowledge-enhanced transform-based multimodal classifier for microbial keratitis identification

Microbial keratitis, a nonviral corneal infection caused by bacteria, fungi, and protozoa, is an urgent condition in ophthalmology requiring prompt treatment in order to prevent severe complications of corneal perforation and vision loss. It is difficult to distinguish between bacterial and fungal keratitis from image unimodal alone, as the characteristics of the sample images themselves are very close. Therefore, this study aims to develop a new deep learning model called knowledge-enhanced transform-based multimodal classifier that exploited the potential of slit-lamp images along with treatment texts to identify bacterial keratitis (BK) and fungal keratitis (FK). The model performance was evaluated in terms of the accuracy, specificity, sensitivity and the area under the curve (AUC). 704 images from 352 patients were divided into training, validation and testing set. In the testing set, our model reached the best accuracy was 93%, sensitivity was 0.97(95% CI [0.84,1]), specificity was 0.92(95% CI [0.76,0.98]) and AUC was 0.94(95% CI [0.92,0.96]), exceeding the benchmark accuracy of 0.86. The diagnostic average accuracies of BK ranged from 81 to 92%, respectively and those for FK were 89-97%. It is the first study to focus on the influence of disease changes and medication interventions on infectious keratitis and our model outperformed the benchmark models and reaching the state-of-the-art performance.

A multi-centre prospective evaluation of THEIA™ to detect diabetic retinopathy (DR) and diabetic macular oedema (DMO) in the New Zealand screening program

PURPOSE

To validate the potential application of THEIA™ as clinical decision making assistant in a national screening program.

METHODS

A total of 900 patients were recruited from either an urban large eye hospital, or a semi-rural optometrist led screening provider, as they were attending their appointment as part of New Zealand Diabetic Eye Screening Programme. The de-identified images were independently graded by three senior specialists, and final results were aggregated using New Zealand grading scheme, which was then converted to referable/non-referable and Healthy/mild/more than mild/sight threatening categories.

RESULTS

THEIA™ managed to grade all images obtained during the study. Comparing the adjudicated images from the specialist grading team, "ground truth", with the grading by the AI platform in detecting "sight threatening" disease, at the patient level THEIA™ achieved 100% imageability, 100% [98.49-100.00%] sensitivity and [97.02-99.16%] specificity, and negative predictive value of 100%. In other words, THEIA™ did not miss any patients with "more than mild" or "sight threatening" disease. The level of agreement between the clinicians and the aggregated results was (k value: 0.9881, 0.9557, and 0.9175), and the level of agreement between THEIA™ and the aggregated labels was (k value: 0.9515).

CONCLUSION

This multi-centre prospective trial showed that THEIA™ did not miss referable disease when screening for diabetic retinopathy and maculopathy. It also had a very high level of granularity in reporting the disease level. As THEIA™ has been tested on a variety of cameras, operating in a range of clinics (rural/urban, ophthalmologist-led\optometrist-led), we believe that it will be a suitable addition to a public diabetic screening program.

Deep Learning-Based Classification of Subtypes of Primary Angle-Closure Disease With Anterior Segment Optical Coherence Tomography

PRCIS

We developed a deep learning-based classifier that can discriminate primary angle closure suspects (PACS), primary angle closure (PAC)/primary angle closure glaucoma (PACG), and also control eyes with open angle with acceptable accuracy.

PURPOSE

To develop a deep learning-based classifier for differentiating subtypes of primary angle closure disease, including PACS and PAC/PACG, and also normal control eyes.

MATERIALS AND METHODS

Anterior segment optical coherence tomography images were used for analysis with 5 different networks including MnasNet, MobileNet, ResNet18, ResNet50, and EfficientNet. The data set was split with randomization performed at the patient level into a training plus validation set (85%), and a test data set (15%). Then 4-fold cross-validation was used to train the model. In each mentioned architecture, the networks were trained with original and cropped images. Also, the analyses were carried out for single images and images grouped on the patient level (case-based). Then majority voting was applied to the determination of the final prediction.

RESULTS

A total of 1616 images of normal eyes (87 eyes), 1055 images of PACS (66 eyes), and 1076 images of PAC/PACG (66 eyes) eyes were included in the analysis. The mean ± SD age was 51.76 ± 15.15 years and 48.3% were males. MobileNet had the best performance in the model, in which both original and cropped images were used. The accuracy of MobileNet for detecting normal, PACS, and PAC/PACG eyes was 0.99 ± 0.00, 0.77 ± 0.02, and 0.77 ± 0.03, respectively. By running MobileNet in a case-based classification approach, the accuracy improved and reached 0.95 ± 0.03, 0.83 ± 0.06, and 0.81 ± 0.05, respectively. For detecting the open angle, PACS, and PAC/PACG, the MobileNet classifier achieved an area under the curve of 1, 0.906, and 0.872, respectively, on the test data set.

CONCLUSION

The MobileNet-based classifier can detect normal, PACS, and PAC/PACG eyes with acceptable accuracy based on anterior segment optical coherence tomography images.

Determinants of late presentation of glaucoma in Hong Kong

BACKGROUND

Glaucoma is the commonest cause of irreversible blindness worldwide. As it is typically asymptomatic until advanced, the risk of blindness from late presentation is higher than other eye diseases. This study aims to investigate the risk factors for late presentation of primary glaucoma patients.

METHODS

We undertook a hospital-based case-control study of a random sample of glaucoma patients from a hospital in Hong Kong. Structured questionnaires and existing information from the electronic patient record were used, and the odds of presenting late were analysed by logistic regression.

RESULTS

Of 210 recruited participants, 83 (39.5%) presented with advanced glaucoma unilaterally or bilaterally. The mean age of participants was 61.1 ± 11.9 years, with 110 males (52.4%). Univariate analysis revealed that male sex and primary angle-closure glaucoma (PACG) have 3.06 (CI95:1.71-5.48; P < 0.001) and 2.47 (CI95:1.11-5.49; P = 0.03) times higher odds of late presentation, respectively. Multivariate analysis revealed late presenters were 3.54 (CI95:1.35-9.35; P = 0.01) times more likely to have PACG than primary open-angle glaucoma (POAG). Patients with elevated baseline intraocular pressure (IOP) also had 1.06 times higher odds of presenting with advanced glaucoma (CI95:1.02-1.11; P = 0.002). Linear regression revealed that PACG patients present with 7.12 mmHg higher IOP than POAG patients (CI95:4.23-10.0; P < 0.001).

CONCLUSION

In conclusion, a high proportion of glaucoma patients present late in Hong Kong, with gender and type of glaucoma being significant determinants. Our study shows that PACG presents with higher IOP and, along with male gender, are more likely to have advanced disease than POAG.

Association of Biomarker-Based Artificial Intelligence With Risk of Racial Bias in Retinal Images

Importance

Although race is a social construct, it is associated with variations in skin and retinal pigmentation. Image-based medical artificial intelligence (AI) algorithms that use images of these organs have the potential to learn features associated with self-reported race (SRR), which increases the risk of racially biased performance in diagnostic tasks; understanding whether this information can be removed, without affecting the performance of AI algorithms, is critical in reducing the risk of racial bias in medical AI.

Objective

To evaluate whether converting color fundus photographs to retinal vessel maps (RVMs) of infants screened for retinopathy of prematurity (ROP) removes the risk for racial bias.

Design, Setting, and Participants

The retinal fundus images (RFIs) of neonates with parent-reported Black or White race were collected for this study. A u-net, a convolutional neural network (CNN) that provides precise segmentation for biomedical images, was used to segment the major arteries and veins in RFIs into grayscale RVMs, which were subsequently thresholded, binarized, and/or skeletonized. CNNs were trained with patients' SRR labels on color RFIs, raw RVMs, and thresholded, binarized, or skeletonized RVMs. Study data were analyzed from July 1 to September 28, 2021.

Main Outcomes and Measures

Area under the precision-recall curve (AUC-PR) and area under the receiver operating characteristic curve (AUROC) at both the image and eye level for classification of SRR.

Results

A total of 4095 RFIs were collected from 245 neonates with parent-reported Black (94 [38.4%]; mean [SD] age, 27.2 [2.3] weeks; 55 majority sex [58.5%]) or White (151 [61.6%]; mean [SD] age, 27.6 [2.3] weeks, 80 majority sex [53.0%]) race. CNNs inferred SRR from RFIs nearly perfectly (image-level AUC-PR, 0.999; 95% CI, 0.999-1.000; infant-level AUC-PR, 1.000; 95% CI, 0.999-1.000). Raw RVMs were nearly as informative as color RFIs (image-level AUC-PR, 0.938; 95% CI, 0.926-0.950; infant-level AUC-PR, 0.995; 95% CI, 0.992-0.998). Ultimately, CNNs were able to learn whether RFIs or RVMs were from Black or White infants regardless of whether images contained color, vessel segmentation brightness differences were nullified, or vessel segmentation widths were uniform.

Conclusions and Relevance

Results of this diagnostic study suggest that it can be very challenging to remove information relevant to SRR from fundus photographs. As a result, AI algorithms trained on fundus photographs have the potential for biased performance in practice, even if based on biomarkers rather than raw images. Regardless of the methodology used for training AI, evaluating performance in relevant subpopulations is critical.

A Global Review of Publicly Available Datasets Containing Fundus Images: Characteristics, Barriers to Access, Usability, and Generalizability

This article provides a comprehensive and up-to-date overview of the repositories that contain color fundus images. We analyzed them regarding availability and legality, presented the datasets' characteristics, and identified labeled and unlabeled image sets. This study aimed to complete all publicly available color fundus image datasets to create a central catalog of available color fundus image datasets.

MSHF: A Multi-Source Heterogeneous Fundus (MSHF) Dataset for Image Quality Assessment

Image quality assessment (IQA) is significant for current techniques of image-based computer-aided diagnosis, and fundus imaging is the chief modality for screening and diagnosing ophthalmic diseases. However, most of the existing IQA datasets are single-center datasets, disregarding the type of imaging device, eye condition, and imaging environment. In this paper, we collected a multi-source heterogeneous fundus (MSHF) database. The MSHF dataset consisted of 1302 high-resolution normal and pathologic images from color fundus photography (CFP), images of healthy volunteers taken with a portable camera, and ultrawide-field (UWF) images of diabetic retinopathy patients. Dataset diversity was visualized with a spatial scatter plot. Image quality was determined by three ophthalmologists according to its illumination, clarity, contrast and overall quality. To the best of our knowledge, this is one of the largest fundus IQA datasets and we believe this work will be beneficial to the construction of a standardized medical image database.

A bibliometric analysis of artificial intelligence applications in macular edema: exploring research hotspots and Frontiers

Background: Artificial intelligence (AI) is used in ophthalmological disease screening and diagnostics, medical image diagnostics, and predicting late-disease progression rates. We reviewed all AI publications associated with macular edema (ME) research Between 2011 and 2022 and performed modeling, quantitative, and qualitative investigations. Methods: On 1st February 2023, we screened the Web of Science Core Collection for AI applications related to ME, from which 297 studies were identified and analyzed (2011-2022). We collected information on: publications, institutions, country/region, keywords, journal name, references, and research hotspots. Literature clustering networks and Frontier knowledge bases were investigated using bibliometrix-BiblioShiny, VOSviewer, and CiteSpace bibliometric platforms. We used the R "bibliometrix" package to synopsize our observations, enumerate keywords, visualize collaboration networks between countries/regions, and generate a topic trends plot. VOSviewer was used to examine cooperation between institutions and identify citation relationships between journals. We used CiteSpace to identify clustering keywords over the timeline and identify keywords with the strongest citation bursts. Results: In total, 47 countries published AI studies related to ME; the United States had the highest H-index, thus the greatest influence. China and the United States cooperated most closely between all countries. Also, 613 institutions generated publications - the Medical University of Vienna had the highest number of studies. This publication record and H-index meant the university was the most influential in the ME field. Reference clusters were also categorized into 10 headings: retinal Optical Coherence Tomography (OCT) fluid detection, convolutional network models, deep learning (DL)-based single-shot predictions, retinal vascular disease, diabetic retinopathy (DR), convolutional neural networks (CNNs), automated macular pathology diagnosis, dry age-related macular degeneration (DARMD), class weight, and advanced DL architecture systems. Frontier keywords were represented by diabetic macular edema (DME) (2021-2022). Conclusion: Our review of the AI-related ME literature was comprehensive, systematic, and objective, and identified future trends and current hotspots. With increased DL outputs, the ME research focus has gradually shifted from manual ME examinations to automatic ME detection and associated symptoms. In this review, we present a comprehensive and dynamic overview of AI in ME and identify future research areas.

Diagnostic performance of deep learning in infectious keratitis: a systematic review and meta-analysis protocol

INTRODUCTION

Infectious keratitis (IK) represents the fifth-leading cause of blindness worldwide. A delay in diagnosis is often a major factor in progression to irreversible visual impairment and/or blindness from IK. The diagnostic challenge is further compounded by low microbiological culture yield, long turnaround time, poorly differentiated clinical features and polymicrobial infections. In recent years, deep learning (DL), a subfield of artificial intelligence, has rapidly emerged as a promising tool in assisting automated medical diagnosis, clinical triage and decision-making, and improving workflow efficiency in healthcare services. Recent studies have demonstrated the potential of using DL in assisting the diagnosis of IK, though the accuracy remains to be elucidated. This systematic review and meta-analysis aims to critically examine and compare the performance of various DL models with clinical experts and/or microbiological results (the current 'gold standard') in diagnosing IK, with an aim to inform practice on the clinical applicability and deployment of DL-assisted diagnostic models.

METHODS AND ANALYSIS

This review will consider studies that included application of any DL models to diagnose patients with suspected IK, encompassing bacterial, fungal, protozoal and/or viral origins. We will search various electronic databases, including EMBASE and MEDLINE, and trial registries. There will be no restriction to the language and publication date. Two independent reviewers will assess the titles, abstracts and full-text articles. Extracted data will include details of each primary studies, including title, year of publication, authors, types of DL models used, populations, sample size, decision threshold and diagnostic performance. We will perform meta-analyses for the included primary studies when there are sufficient similarities in outcome reporting.

ETHICS AND DISSEMINATION

No ethical approval is required for this systematic review. We plan to disseminate our findings via presentation/publication in a peer-reviewed journal.

PROSPERO REGISTRATION NUMBER

CRD42022348596.

Real-time estimation of the remaining surgery duration for cataract surgery using deep convolutional neural networks and long short-term memory

PURPOSE

Estimating the surgery length has the potential to be utilized as skill assessment, surgical training, or efficient surgical facility utilization especially if it is done in real-time as a remaining surgery duration (RSD). Surgical length reflects a certain level of efficiency and mastery of the surgeon in a well-standardized surgery such as cataract surgery. In this paper, we design and develop a real-time RSD estimation method for cataract surgery that does not require manual labeling and is transferable with minimum fine-tuning.

METHODS

A regression method consisting of convolutional neural networks (CNNs) and long short-term memory (LSTM) is designed for RSD estimation. The model is firstly trained and evaluated for the single main surgeon with a large number of surgeries. Then, the fine-tuning strategy is used to transfer the model to the data of the other two surgeons. Mean Absolute Error (MAE in seconds) was used to evaluate the performance of the RSD estimation. The proposed method is compared with the naïve method which is based on the statistic of the historical data. A transferability experiment is also set to demonstrate the generalizability of the method.

RESULT

The mean surgical time for the sample videos was 318.7 s (s) (standard deviation 83.4 s) for the main surgeon for the initial training. In our experiments, the lowest MAE of 19.4 s (equal to about 6.4% of the mean surgical time) is achieved by our best-trained model for the independent test data of the main target surgeon. It reduces the MAE by 35.5 s (-10.2%) compared to the naïve method. The fine-tuning strategy transfers the model trained for the main target to the data of other surgeons with only a small number of training data (20% of the pre-training). The MAEs for the other two surgeons are 28.3 s and 30.6 s with the fine-tuning model, which decreased by -8.1 s and -7.5 s than the Per-surgeon model (average declining of -7.8 s and 1.3% of video duration). External validation study with Cataract-101 outperformed 3 reported methods of TimeLSTM, RSDNet, and CataNet.

CONCLUSION

An approach to build a pre-trained model for estimating RSD estimation based on a single surgeon and then transfer to other surgeons demonstrated both low prediction error and good transferability with minimum fine-tuning videos.

Serum cholinesterase is associated with incident diabetic retinopathy: the Shanghai Nicheng cohort study

BACKGROUND

Serum cholinesterase (ChE) is positively associated with incident diabetes and dyslipidemia. We aimed to investigate the relationship between ChE and the incidence of diabetic retinopathy (DR).

METHODS

Based on a community-based cohort study followed for 4.6 years, 1133 participants aged 55-70 years with diabetes were analyzed. Fundus photographs were taken for each eye at both baseline and follow-up investigations. The presence and severity of DR were categorized into no DR, mild non-proliferative DR (NPDR), and referable DR (moderate NPDR or worse). Binary and multinomial logistic regression models were used to estimate the risk ratio (RR) and 95% confidence interval (CI) between ChE and DR.

RESULTS

Among the 1133 participants, 72 (6.4%) cases of DR occurred. The multivariable binary logistic regression showed that the highest tertile of ChE (≥ 422 U/L) was associated with a 2.01-fold higher risk of incident DR (RR 2.01, 95%CI 1.01-4.00; P for trend < 0.05) than the lowest tertile (< 354 U/L). The multivariable binary and multinomial logistic regression showed that the risk of DR increased by 41% (RR 1.41, 95%CI 1.05-1.90), and the risk of incident referable DR was almost 2-fold higher than no DR (RR 1.99, 95%CI 1.24-3.18) with per 1-SD increase of log_e-transformed ChE. Furthermore, multiplicative interactions were found between ChE and elderly participants (aged 60 and older; P for interaction = 0.003) and men (P for interaction = 0.044) on the risk of DR.

CONCLUSIONS

In this study, ChE was associated with the incidence of DR, especially referable DR. ChE was a potential biomarker for predicting the incident DR.

Multicentre external validation of a commercial artificial intelligence software to analyse chest radiographs in health screening environments with low disease prevalence

OBJECTIVES

To externally validate the performance of a commercial AI software program for interpreting CXRs in a large, consecutive, real-world cohort from primary healthcare centres.

METHODS

A total of 3047 CXRs were collected from two primary healthcare centres, characterised by low disease prevalence, between January and December 2018. All CXRs were labelled as normal or abnormal according to CT findings. Four radiology residents read all CXRs twice with and without AI assistance. The performances of the AI and readers with and without AI assistance were measured in terms of area under the receiver operating characteristic curve (AUROC), sensitivity, and specificity.

RESULTS

The prevalence of clinically significant lesions was 2.2% (68 of 3047). The AUROC, sensitivity, and specificity of the AI were 0.648 (95% confidence interval [CI] 0.630-0.665), 35.3% (CI, 24.7-47.8), and 94.2% (CI, 93.3-95.0), respectively. AI detected 12 of 41 pneumonia, 3 of 5 tuberculosis, and 9 of 22 tumours. AI-undetected lesions tended to be smaller than true-positive lesions. The readers' AUROCs ranged from 0.534-0.676 without AI and 0.571-0.688 with AI (all p values < 0.05). For all readers, the mean reading time was 2.96-10.27 s longer with AI assistance (all p values < 0.05).

CONCLUSIONS

The performance of commercial AI in these high-volume, low-prevalence settings was poorer than expected, although it modestly boosted the performance of less-experienced readers. The technical prowess of AI demonstrated in experimental settings and approved by regulatory bodies may not directly translate to real-world practice, especially where the demand for AI assistance is highest.

KEY POINTS

• This study shows the limited applicability of commercial AI software for detecting abnormalities in CXRs in a health screening population. • When using AI software in a specific clinical setting that differs from the training setting, it is necessary to adjust the threshold or perform additional training with such data that reflects this environment well. • Prospective test accuracy studies, randomised controlled trials, or cohort studies are needed to examine AI software to be implemented in real clinical practice.

PREDICTING VISUAL OUTCOME AFTER SURGERY IN PATIENTS WITH IDIOPATHIC EPIRETINAL MEMBRANE USING A NOVEL CONVOLUTIONAL NEURAL NETWORK

PURPOSE

To develop a deep convolutional neural network that enables the prediction of postoperative visual outcomes after epiretinal membrane surgery based on preoperative optical coherence tomography images and clinical parameters to refine surgical decision making.

METHODS

A total of 529 patients with idiopathic epiretinal membrane who underwent standard vitrectomy with epiretinal membrane peeling surgery by two surgeons between January 1, 2014, and June 1, 2020, were enrolled. The newly developed Heterogeneous Data Fusion Net was introduced to predict postoperative visual acuity outcomes (improvement ≥2 lines in Snellen chart) 12 months after surgery based on preoperative cross-sectional optical coherence tomography images and clinical factors, including age, sex, and preoperative visual acuity. The predictive accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve of the convolutional neural network model were evaluated.

RESULTS

The developed model demonstrated an overall accuracy for visual outcome prediction of 88.68% (95% CI, 79.0%-95.7%) with an area under the receiver operating characteristic curve of 97.8% (95% CI, 86.8%-98.0%), sensitivity of 87.0% (95% CI, 67.9%-95.5%), specificity of 92.9% (95% CI, 77.4%-98.0%), precision of 0.909, recall of 0.870, and F1 score of 0.889. The heatmaps identified the critical area for prediction as the ellipsoid zone of photoreceptors and the superficial retina, which was subjected to tangential traction of the proliferative membrane.

CONCLUSION

The novel Heterogeneous Data Fusion Net demonstrated high accuracy in the automated prediction of visual outcomes after weighing and leveraging multiple clinical parameters, including optical coherence tomography images. This approach may be helpful in establishing personalized therapeutic strategies for epiretinal membrane management.

Artificial Intelligence in Clinical Oncology: From Data to Digital Pathology and Treatment

Recently, a wide spectrum of artificial intelligence (AI)-based applications in the broader categories of digital pathology, biomarker development, and treatment have been explored. In the domain of digital pathology, these have included novel analytical strategies for realizing new information derived from standard histology to guide treatment selection and biomarker development to predict treatment selection and response. In therapeutics, these have included AI-driven drug target discovery, drug design and repurposing, combination regimen optimization, modulated dosing, and beyond. Given the continued advances that are emerging, it is important to develop workflows that seamlessly combine the various segments of AI innovation to comprehensively augment the diagnostic and interventional arsenal of the clinical oncology community. To overcome challenges that remain with regard to the ideation, validation, and deployment of AI in clinical oncology, recommendations toward bringing this workflow to fruition are also provided from clinical, engineering, implementation, and health care economics considerations. Ultimately, this work proposes frameworks that can potentially integrate these domains toward the sustainable adoption of practice-changing AI by the clinical oncology community to drive improved patient outcomes.

The three-dimensional weakly supervised deep learning algorithm for traumatic splenic injury detection and sequential localization: an experimental study

BACKGROUND

Splenic injury is the most common solid visceral injury in blunt abdominal trauma, and high-resolution abdominal computed tomography (CT) can adequately detect the injury. However, these lethal injuries sometimes have been overlooked in current practice. Deep learning (DL) algorithms have proven their capabilities in detecting abnormal findings in medical images. The aim of this study is to develop a three-dimensional, weakly supervised DL algorithm for detecting splenic injury on abdominal CT using a sequential localization and classification approach.

MATERIAL AND METHODS

The dataset was collected in a tertiary trauma center on 600 patients who underwent abdominal CT between 2008 and 2018, half of whom had splenic injuries. The images were split into development and test datasets at a 4 : 1 ratio. A two-step DL algorithm, including localization and classification models, was constructed to identify the splenic injury. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Grad-CAM (Gradient-weighted Class Activation Mapping) heatmaps from the test set were visually assessed. To validate the algorithm, we also collected images from another hospital to serve as external validation data.

RESULTS

A total of 480 patients, 50% of whom had spleen injuries, were included in the development dataset, and the rest were included in the test dataset. All patients underwent contrast-enhanced abdominal CT in the emergency room. The automatic two-step EfficientNet model detected splenic injury with an AUROC of 0.901 (95% CI: 0.836-0.953). At the maximum Youden index, the accuracy, sensitivity, specificity, PPV, and NPV were 0.88, 0.81, 0.92, 0.91, and 0.83, respectively. The heatmap identified 96.3% of splenic injury sites in true positive cases. The algorithm achieved a sensitivity of 0.92 for detecting trauma in the external validation cohort, with an acceptable accuracy of 0.80.

CONCLUSIONS

The DL model can identify splenic injury on CT, and further application in trauma scenarios is possible.

Artificial intelligence innovation in healthcare: Relevance of reporting guidelines for clinical translation from bench to bedside

Artificial intelligence (AI) and digital innovation are transforming healthcare. Technologies such as machine learning in image analysis, natural language processing in medical chatbots and electronic medical record extraction have the potential to improve screening, diagnostics and prognostication, leading to precision medicine and preventive health. However, it is crucial to ensure that AI research is conducted with scientific rigour to facilitate clinical implementation. Therefore, reporting guidelines have been developed to standardise and streamline the development and validation of AI technologies in health. This commentary proposes a structured approach to utilise these reporting guidelines for the translation of promising AI techniques from research and development into clinical translation, and eventual widespread implementation from bench to bedside.

Contrastive learning-based pretraining improves representation and transferability of diabetic retinopathy classification models

Diabetic retinopathy (DR) is a major cause of vision impairment in diabetic patients worldwide. Due to its prevalence, early clinical diagnosis is essential to improve treatment management of DR patients. Despite recent demonstration of successful machine learning (ML) models for automated DR detection, there is a significant clinical need for robust models that can be trained with smaller cohorts of dataset and still perform with high diagnostic accuracy in independent clinical datasets (i.e., high model generalizability). Towards this need, we have developed a self-supervised contrastive learning (CL) based pipeline for classification of referable vs non-referable DR. Self-supervised CL based pretraining allows enhanced data representation, therefore, the development of robust and generalized deep learning (DL) models, even with small, labeled datasets. We have integrated a neural style transfer (NST) augmentation in the CL pipeline to produce models with better representations and initializations for the detection of DR in color fundus images. We compare our CL pretrained model performance with two state of the art baseline models pretrained with Imagenet weights. We further investigate the model performance with reduced labeled training data (down to 10 percent) to test the robustness of the model when trained with small, labeled datasets. The model is trained and validated on the EyePACS dataset and tested independently on clinical datasets from the University of Illinois, Chicago (UIC). Compared to baseline models, our CL pretrained FundusNet model had higher area under the receiver operating characteristics (ROC) curve (AUC) (CI) values (0.91 (0.898 to 0.930) vs 0.80 (0.783 to 0.820) and 0.83 (0.801 to 0.853) on UIC data). At 10 percent labeled training data, the FundusNet AUC was 0.81 (0.78 to 0.84) vs 0.58 (0.56 to 0.64) and 0.63 (0.60 to 0.66) in baseline models, when tested on the UIC dataset. CL based pretraining with NST significantly improves DL classification performance, helps the model generalize well (transferable from EyePACS to UIC data), and allows training with small, annotated datasets, therefore reducing ground truth annotation burden of the clinicians.

Global trends and performances in diabetic retinopathy studies: A bibliometric analysis

Objective

The objective of this study is to conduct a comprehensive bibliometric analysis to identify and evaluate global trends in diabetic retinopathy (DR) research and visualize the focus and frontiers of this field.

Methods

Diabetic retinopathy-related publications from the establishment of the Web of Science (WOS) through 1 November 2022 were retrieved for qualitative and quantitative analyses. This study analyzed annual publication counts, prolific countries, institutions, journals, and the top 10 most cited literature. The findings were presented through descriptive statistics. VOSviewer 1.6.17 was used to exhibit keywords with high frequency and national cooperation networks, while CiteSpace 5.5.R2 displayed the timeline and burst keywords for each term.

Results

A total of 10,709 references were analyzed, and the number of publications continuously increased over the investigated period. America had the highest h-index and citation frequency, contributing to the most influence. China was the most prolific country, producing 3,168 articles. The University of London had the highest productivity. The top three productive journals were from America, and Investigative Ophthalmology Visual Science had the highest number of publications. The article from Gulshan et al. (2016; co-citation counts, 2,897) served as the representative and symbolic reference. The main research topics in this area were incidence, pathogenesis, treatment, and artificial intelligence (AI). Deep learning, models, biomarkers, and optical coherence tomography angiography (OCTA) of DR were frontier hotspots.

Conclusion

Bibliometric analysis in this study provided valuable insights into global trends in DR research frontiers. Four key study directions and three research frontiers were extracted from the extensive DR-related literature. As the incidence of DR continues to increase, DR prevention and treatment have become a pressing public health concern and a significant area of research interest. In addition, the development of AI technologies and telemedicine has emerged as promising research frontiers for balancing the number of doctors and patients.

An interpretable and interactive deep learning algorithm for a clinically applicable retinal fundus diagnosis system by modelling finding-disease relationship

The identification of abnormal findings manifested in retinal fundus images and diagnosis of ophthalmic diseases are essential to the management of potentially vision-threatening eye conditions. Recently, deep learning-based computer-aided diagnosis systems (CADs) have demonstrated their potential to reduce reading time and discrepancy amongst readers. However, the obscure reasoning of deep neural networks (DNNs) has been the leading cause to reluctance in its clinical use as CAD systems. Here, we present a novel architectural and algorithmic design of DNNs to comprehensively identify 15 abnormal retinal findings and diagnose 8 major ophthalmic diseases from macula-centered fundus images with the accuracy comparable to experts. We then define a notion of counterfactual attribution ratio (CAR) which luminates the system's diagnostic reasoning, representing how each abnormal finding contributed to its diagnostic prediction. By using CAR, we show that both quantitative and qualitative interpretation and interactive adjustment of the CAD result can be achieved. A comparison of the model's CAR with experts' finding-disease diagnosis correlation confirms that the proposed model identifies the relationship between findings and diseases similarly as ophthalmologists do.

Artificial intelligence to estimate the tear film breakup time and diagnose dry eye disease

The use of artificial intelligence (AI) in the diagnosis of dry eye disease (DED) remains limited due to the lack of standardized image formats and analysis models. To overcome these issues, we used the Smart Eye Camera (SEC), a video-recordable slit-lamp device, and collected videos of the anterior segment of the eye. This study aimed to evaluate the accuracy of the AI algorithm in estimating the tear film breakup time and apply this model for the diagnosis of DED according to the Asia Dry Eye Society (ADES) DED diagnostic criteria. Using the retrospectively corrected DED videos of 158 eyes from 79 patients, 22,172 frames were annotated by the DED specialist to label whether or not the frame had breakup. The AI algorithm was developed using the training dataset and machine learning. The DED criteria of the ADES was used to determine the diagnostic performance. The accuracy of tear film breakup time estimation was 0.789 (95% confidence interval (CI) 0.769-0.809), and the area under the receiver operating characteristic curve of this AI model was 0.877 (95% CI 0.861-0.893). The sensitivity and specificity of this AI model for the diagnosis of DED was 0.778 (95% CI 0.572-0.912) and 0.857 (95% CI 0.564-0.866), respectively. We successfully developed a novel AI-based diagnostic model for DED. Our diagnostic model has the potential to enable ophthalmology examination outside hospitals and clinics.

Application of big data in ophthalmology

The advents of information technologies have led to the creation of ever-larger datasets. Also known as big data, these large datasets are characterized by its volume, variety, velocity, veracity, and value. More importantly, big data has the potential to expand traditional research capabilities, inform clinical practice based on real-world data, and improve the health system and service delivery. This review first identified the different sources of big data in ophthalmology, including electronic medical records, data registries, research consortia, administrative databases, and biobanks. Then, we provided an in-depth look at how big data analytics have been applied in ophthalmology for disease surveillance, and evaluation on disease associations, detection, management, and prognostication. Finally, we discussed the challenges involved in big data analytics, such as data suitability and quality, data security, and analytical methodologies.

Using Artificial Intelligence to Analyse the Retinal Vascular Network: The Future of Cardiovascular Risk Assessment Based on Oculomics? A Narrative Review

The healthcare burden of cardiovascular diseases remains a major issue worldwide. Understanding the underlying mechanisms and improving identification of people with a higher risk profile of systemic vascular disease through noninvasive examinations is crucial. In ophthalmology, retinal vascular network imaging is simple and noninvasive and can provide in vivo information of the microstructure and vascular health. For more than 10 years, different research teams have been working on developing software to enable automatic analysis of the retinal vascular network from different imaging techniques (retinal fundus photographs, OCT angiography, adaptive optics, etc.) and to provide a description of the geometric characteristics of its arterial and venous components. Thus, the structure of retinal vessels could be considered a witness of the systemic vascular status. A new approach called "oculomics" using retinal image datasets and artificial intelligence algorithms recently increased the interest in retinal microvascular biomarkers. Despite the large volume of associated research, the role of retinal biomarkers in the screening, monitoring, or prediction of systemic vascular disease remains uncertain. A PubMed search was conducted until August 2022 and yielded relevant peer-reviewed articles based on a set of inclusion criteria. This literature review is intended to summarize the state of the art in oculomics and cardiovascular disease research.

TMM-Nets: Transferred Multi- to Mono-Modal Generation for Lupus Retinopathy Diagnosis

Rare diseases, which are severely underrepresented in basic and clinical research, can particularly benefit from machine learning techniques. However, current learning-based approaches usually focus on either mono-modal image data or matched multi-modal data, whereas the diagnosis of rare diseases necessitates the aggregation of unstructured and unmatched multi-modal image data due to their rare and diverse nature. In this study, we therefore propose diagnosis-guided multi-to-mono modal generation networks (TMM-Nets) along with training and testing procedures. TMM-Nets can transfer data from multiple sources to a single modality for diagnostic data structurization. To demonstrate their potential in the context of rare diseases, TMM-Nets were deployed to diagnose the lupus retinopathy (LR-SLE), leveraging unmatched regular and ultra-wide-field fundus images for transfer learning. The TMM-Nets encoded the transfer learning from diabetic retinopathy to LR-SLE based on the similarity of the fundus lesions. In addition, a lesion-aware multi-scale attention mechanism was developed for clinical alerts, enabling TMM-Nets not only to inform patient care, but also to provide insights consistent with those of clinicians. An adversarial strategy was also developed to refine multi- to mono-modal image generation based on diagnostic results and the data distribution to enhance the data augmentation performance. Compared to the baseline model, the TMM-Nets showed 35.19% and 33.56% F1 score improvements on the test and external validation sets, respectively. In addition, the TMM-Nets can be used to develop diagnostic models for other rare diseases.

Artificial intelligence in diabetic retinopathy: Bibliometric analysis

BACKGROUND

The use of artificial intelligence in diabetic retinopathy has become a popular research focus in the past decade. However, no scientometric report has provided a systematic overview of this scientific area.

AIMS

We utilized a bibliometric approach to identify and analyse the academic literature on artificial intelligence in diabetic retinopathy and explore emerging research trends, key authors, co-authorship networks, institutions, countries, and journals. We further captured the diabetic retinopathy conditions and technology commonly used within this area.

METHODS

Web of Science was used to collect relevant articles on artificial intelligence use in diabetic retinopathy published between January 1, 2012, and December 31, 2022 . All the retrieved titles were screened for eligibility, with one criterion that they must be in English. All the bibliographic information was extracted and used to perform a descriptive analysis. Bibliometrix (R tool) and VOSviewer (Leiden University) were used to construct and visualize the annual numbers of publications, journals, authors, countries, institutions, collaboration networks, keywords, and references.

RESULTS

In total, 931 articles that met the criteria were collected. The number of annual publications showed an increasing trend over the last ten years. Investigative Ophthalmology & Visual Science (58/931), IEEE Access (54/931), and Computers in Biology and Medicine (23/931) were the most journals with most publications. China (211/931), India (143/931, USA (133/931), and South Korea (44/931) were the most productive countries of origin. The National University of Singapore (40/931), Singapore Eye Research Institute (35/931), and Johns Hopkins University (34/931) were the most productive institutions. Ting D. (34/931), Wong T. (28/931), and Tan G. (17/931) were the most productive researchers.

CONCLUSION

This study summarizes the recent advances in artificial intelligence technology on diabetic retinopathy research and sheds light on the emerging trends, sources, leading institutions, and hot topics through bibliometric analysis and network visualization. Although this field has already shown great potential in health care, our findings will provide valuable clues relevant to future research directions and clinical practice.

Generalisability and performance of an OCT-based deep learning classifier for community-based and hospital-based detection of gonioscopic angle closure

PURPOSE

To assess the generalisability and performance of a deep learning classifier for automated detection of gonioscopic angle closure in anterior segment optical coherence tomography (AS-OCT) images.

METHODS

A convolutional neural network (CNN) model developed using data from the Chinese American Eye Study (CHES) was used to detect gonioscopic angle closure in AS-OCT images with reference gonioscopy grades provided by trained ophthalmologists. Independent test data were derived from the population-based CHES, a community-based clinic in Singapore, and a hospital-based clinic at the University of Southern California (USC). Classifier performance was evaluated with receiver operating characteristic curve and area under the receiver operating characteristic curve (AUC) metrics. Interexaminer agreement between the classifier and two human examiners at USC was calculated using Cohen's kappa coefficients.

RESULTS

The classifier was tested using 640 images (311 open and 329 closed) from 127 Chinese Americans, 10 165 images (9595 open and 570 closed) from 1318 predominantly Chinese Singaporeans and 300 images (234 open and 66 closed) from 40 multiethnic USC patients. The classifier achieved similar performance in the CHES (AUC=0.917), Singapore (AUC=0.894) and USC (AUC=0.922) cohorts. Standardising the distribution of gonioscopy grades across cohorts produced similar AUC metrics (range 0.890-0.932). The agreement between the CNN classifier and two human examiners (Ҡ=0.700 and 0.704) approximated interexaminer agreement (Ҡ=0.693) in the USC cohort.

CONCLUSION

An OCT-based deep learning classifier demonstrated consistent performance detecting gonioscopic angle closure across three independent patient populations. This automated method could aid ophthalmologists in the assessment of angle status in diverse patient populations.

Eyes as the windows into cardiovascular disease in the era of big data

Cardiovascular disease (CVD) is a major cause of mortality and morbidity worldwide and imposes significant socioeconomic burdens, especially with late diagnoses. There is growing evidence of strong correlations between ocular images, which are information-dense, and CVD progression. The accelerating development of deep learning algorithms (DLAs) is a promising avenue for research into CVD biomarker discovery, early CVD diagnosis, and CVD prognostication. We review a selection of 17 recent DLAs on the less-explored realm of DL as applied to ocular images to produce CVD outcomes, potential challenges in their clinical deployment, and the path forward. The evidence for CVD manifestations in ocular images is well documented. Most of the reviewed DLAs analyze retinal fundus photographs to predict CV risk factors, in particular hypertension. DLAs can predict age, sex, smoking status, alcohol status, body mass index, mortality, myocardial infarction, stroke, chronic kidney disease, and hematological disease with significant accuracy. While the cardio-oculomics intersection is now burgeoning, very much remain to be explored. The increasing availability of big data, computational power, technological literacy, and acceptance all prime this subfield for rapid growth. We pinpoint the specific areas of improvement toward ubiquitous clinical deployment: increased generalizability, external validation, and universal benchmarking. DLAs capable of predicting CVD outcomes from ocular inputs are of great interest and promise to individualized precision medicine and efficiency in the provision of health care with yet undetermined real-world efficacy with impactful initial results.

Extensive deep learning model to enhance electrocardiogram application via latent cardiovascular feature extraction from identity identification

BACKGROUND AND OBJECTIVE

Deep learning models (DLMs) have been successfully applied in biomedicine primarily using supervised learning with large, annotated databases. However, scarce training resources limit the potential of DLMs for electrocardiogram (ECG) analysis.

METHODS

We have developed a novel pre-training strategy for unsupervised identity identification with an area under the receiver operating characteristic curve (AUC) >0.98. Accordingly, a DLM pre-trained with identity identification can be applied to 70 patient characteristic predictions using transfer learning (TL). These ECG-based patient characteristics were then used for cardiovascular disease (CVD) risk prediction. The DLMs were trained using 507,729 ECGs from 222,473 patients and validated using two independent validation sets (n = 27,824/31,925).

RESULTS

The DLMs using our method exhibited better performance than directly trained DLMs. Additionally, our DLM performed better than those of previous studies in terms of gender (AUC [internal/external] = 0.982/0.968), age (correlation = 0.886/0.892), low ejection fraction (AUC = 0.942/0.951), and critical markers not addressed previously, including high B-type natriuretic peptide (AUC = 0.921/0.899). Additionally, approximately 50% of the ECG-based characteristics provided significantly more prediction information for cardiovascular risk than real characteristics.

CONCLUSIONS

This is the first study to use identity identification as a pre-training task for TL in ECG analysis. An extensive exploration of the relationship between ECG and 70 patient characteristics was conducted. Our DLM-enhanced ECG interpretation system extensively advanced ECG-related patient characteristic prediction and mortality risk management for cardiovascular diseases.

A Primer in Precision Nephrology: Optimizing Outcomes in Kidney Health and Disease through Data-Driven Medicine

This year marks the 63rd anniversary of the International Society of Nephrology, which signaled nephrology's emergence as a modern medical discipline. In this article, we briefly trace the course of nephrology's history to show a clear arc in its evolution-of increasing resolution in nephrological data-an arc that is converging with computational capabilities to enable precision nephrology. In general, precision medicine refers to tailoring treatment to the individual characteristics of patients. For an operational definition, this tailoring takes the form of an optimization, in which treatments are selected to maximize a patient's expected health with respect to all available data. Because modern health data are large and high resolution, this optimization process requires computational intervention, and it must be tuned to the contours of specific medical disciplines. An advantage of this operational definition for precision medicine is that it allows us to better understand what precision medicine means in the context of a specific medical discipline. The goal of this article was to demonstrate how to instantiate this definition of precision medicine for the field of nephrology. Correspondingly, the goal of precision nephrology was to answer two related questions: ( 1 ) How do we optimize kidney health with respect to all available data? and ( 2 ) How do we optimize general health with respect to kidney data?

Development and validation of a risk prediction model for diabetic retinopathy in type 2 diabetic patients

To establish a risk prediction model and make individualized assessment for the susceptible diabetic retinopathy (DR) population in type 2 diabetic mellitus (T2DM) patients. According to the retrieval strategy, inclusion and exclusion criteria, the relevant meta-analyses on DR risk factors were searched and evaluated. The pooled odds ratio (OR) or relative risk (RR) of each risk factor was obtained and calculated for β coefficients using logistic regression (LR) model. Besides, an electronic patient-reported outcome questionnaire was developed and 60 cases of DR and non-DR T2DM patients were investigated to validate the developed model. Receiver operating characteristic curve (ROC) was drawn to verify the prediction accuracy of the model. After retrieving, eight meta-analyses with a total of 15,654 cases and 12 risk factors associated with the onset of DR in T2DM, including weight loss surgery, myopia, lipid-lowing drugs, intensive glucose control, course of T2DM, glycated hemoglobin (HbA1c), fasting plasma glucose, hypertension, gender, insulin treatment, residence, and smoking were included for LR modeling. These factors, followed by the respective β coefficient was bariatric surgery (− 0.942), myopia (− 0.357), lipid-lowering drug follow-up < 3y (− 0.994), lipid-lowering drug follow-up > 3y (− 0.223), course of T2DM (0.174), HbA1c (0.372), fasting plasma glucose (0.223), insulin therapy (0.688), rural residence (0.199), smoking (− 0.083), hypertension (0.405), male (0.548), intensive glycemic control (− 0.400) with constant term α (− 0.949) in the constructed model. The area under receiver operating characteristic curve (AUC) of the model in the external validation was 0.912. An application was presented as an example of use. In conclusion, the risk prediction model of DR is developed, which makes individualized assessment for the susceptible DR population feasible and needs to be further verified with large sample size application.

Development and validation of a deep learning model using convolutional neural networks to identify femoral internal fixation device in radiographs

OBJECTIVE

The purpose of this study is to develop and validate a deep convolutional neural network (DCNN) model to automatically identify the manufacturer and model of hip internal fixation devices from anteroposterior (AP) radiographs.

MATERIALS AND METHODS

In this retrospective study, 1721 hip AP radiographs, including six internal fixation devices from 1012 patients, were collected from an orthopedic center between June 2014 and June 2022 to establish a classification network. The images were divided into training set (1106 images), validation set (272 images), and test set (343 images). The model efficacy is evaluated by using the data on the test set. The overall TOP-1 accuracy, and the precision, sensitivity, specificity, and F1 score of each model are calculated, and receiver operating characteristic (ROC) curves are plotted to evaluate the model performance. Gradient-weighted class activation mapping (Grad-CAM) images are used to determine the image features that are most important for DCNN decisions.

RESULTS

A total of 1378 (80%) images were used for model development, and model efficacy was validated on a test set with 343 (20%) images. The overall TOP-1 accuracy was 98.5%. The area under the receiver operating characteristic curve (AUC) values for each internal fixation model were 1.000, 1.000, 0.980, 1.000, 0.999, and 1.000, respectively. Gradient-weighted class activation mapping showed the unique design of the internal fixation device.

CONCLUSION

We developed a deep convolutional neural network model that can identify the manufacturer and model of hip internal fixation devices from the hip AP radiographs.