An explainable deep-learning algorithm for the detection of acute intracranial haemorrhage from small datasets

Deep learning of pretreatment multiphase CT images for predicting response to lenvatinib and immune checkpoint inhibitors in unresectable hepatocellular carcinoma

Objectives

Combination therapy of lenvatinib and immune checkpoint inhibitors (CLICI) has emerged as a promising approach for managing unresectable hepatocellular carcinoma (HCC). However, the response to such treatment is observed in only a subset of patients, underscoring the pressing need for reliable methods to identify potential responders.

Materials & methods

This was a retrospective analysis involving 120 patients with unresectable HCC. They were divided into training (n = 72) and validation (n = 48) cohorts. We developed an interpretable deep learning model using multiphase computed tomography (CT) images to predict whether patients will respond or not to CLICI treatment, based on the Response Evaluation Criteria in Solid Tumors, version 1.1 (RECIST v1.1). We evaluated the models' performance and analyzed the impact of each CT phase. Critical regions influencing predictions were identified and visualized through heatmaps.

Results

The multiphase model outperformed the best biphase and uniphase models, achieving an area under the curve (AUC) of 0.802 (95% CI = 0.780-0.824). The portal phase images were found to significantly enhance the model's predictive accuracy. Heatmaps identified six critical features influencing treatment response, offering valuable insights to clinicians. Additionally, we have made this model accessible via a web server at http://uhccnet.com/ for ease of use.

Conclusions

The integration of multiphase CT images with deep learning-generated heatmaps for predicting treatment response provides a robust and practical tool for guiding CLICI therapy in patients with unresectable HCC.

Applications of artificial intelligence for adolescent idiopathic scoliosis: mapping the evidence

PURPOSE

Adolescent idiopathic scoliosis (AIS) is a common spinal deformity with varying progression, complicating treatment decisions. Artificial intelligence (AI) and machine learning (ML) are increasingly prominent in orthopedic care, aiding in diagnosis, risk-stratification, and treatment guidance. This scoping review outlines AI applications in AIS.

METHODS

This study followed PRISMA-ScR guidelines and included articles that reported the development, use, or validation of AI models for treating, diagnosing, or predicting clinical outcomes in AIS.

RESULTS

40 full-text articles were included, with most studies published in the last 5 years (77.5%). Common ML techniques were convolutional neural networks (55%), decision trees and random forests (15%), and artificial neural networks (15%). Most AI applications in AIS were for imaging analysis (25/40; 62.5%), focusing on automatic measurement of Cobb angle, and axial vertebral rotation (13/25; 52%) and curve classification/severity (13/25; 52%). Prediction was the second most common application (15/40; 37.5%), with studies predicting curve progression (9/15; 60%), and Cobb angles (9/15; 60%). Only 15 studies (37.5%) reported clinical implementation guidelines for AI in AIS management. 52.5% of studies reported model accuracy, with an average of 85.4%.

CONCLUSION

This review highlights the applications of AI in AIS care, notably including automatic radiographic analysis, curve type classification, prediction of curve progression, and AIS diagnosis. However, the current lack of clear clinical implementation guidelines, model transparency, and external validation of studied models limits clinician trust and the generalizability and applicability of AI in AIS management.

Artificial intelligence enhances the management of esophageal squamous cell carcinoma in the precision oncology era

Esophageal squamous cell carcinoma (ESCC) is the most common histological type of esophageal cancer with a poor prognosis. Early diagnosis and prognosis assessment are crucial for improving the survival rate of ESCC patients. With the advancement of artificial intelligence (AI) technology and the proliferation of medical digital information, AI has demonstrated promising sensitivity and accuracy in assisting precise detection, treatment decision-making, and prognosis assessment of ESCC. It has become a unique opportunity to enhance comprehensive clinical management of ESCC in the era of precision oncology. This review examines how AI is applied to the diagnosis, treatment, and prognosis assessment of ESCC in the era of precision oncology, and analyzes the challenges and potential opportunities that AI faces in clinical translation. Through insights into future prospects, it is hoped that this review will contribute to the real-world application of AI in future clinical settings, ultimately alleviating the disease burden caused by ESCC.

A Fully Automated Pipeline Using Swin Transformers for Deep Learning-Based Blood Segmentation on Head Computed Tomography Scans After Aneurysmal Subarachnoid Hemorrhage

BACKGROUND

Accurate volumetric assessment of spontaneous aneurysmal subarachnoid hemorrhage (aSAH) is a labor-intensive task performed with current manual and semiautomatic methods that might be relevant for its clinical and prognostic implications. In the present research, we sought to develop and validate an artificial intelligence-driven, fully automated blood segmentation tool for subarachnoid hemorrhage (SAH) patients via noncontrast computed tomography (NCCT) scans employing a transformer-based Swin-UNETR architecture.

METHODS

We retrospectively analyzed NCCT scans from patients with confirmed aSAH utilizing the Swin-UNETR for segmentation. The performance of the proposed method was evaluated against manually segmented ground truth data using metrics such as Dice score, intersection over union, volumetric similarity index , symmetric average surface distance , sensitivity, and specificity. A validation cohort from an external institution was included to test the generalizability of the model.

RESULTS

The model demonstrated high accuracy with robust performance metrics across the internal and external validation cohorts. Notably, it achieved high Dice coefficient (0.873 ± 0.097), intersection over union (0.810 ± 0.092), volumetric similarity index (0.840 ± 0.131), sensitivity (0.821 ± 0.217), and specificity (0.996 ± 0.004) values and a low symmetric average surface distance (1.866 ± 2.910), suggesting proficiency in segmenting blood in SAH patients. The model's efficiency was reflected in its processing speed, indicating potential for real-time applications.

CONCLUSIONS

Our Swin UNETR-based model offers significant advances in the automated segmentation of blood in SAH patients on NCCT images. Despite the computational demands, the model operates effectively on standard hardware with a user-friendly interface, facilitating broader clinical adoption. Further validation across diverse datasets is warranted to confirm its clinical reliability.

Prospective Evaluation of Artificial Intelligence Triage of Intracranial Hemorrhage on Noncontrast Head CT Examinations

Background: Retrospective studies evaluating artificial intelligence (AI) algorithms for intracranial hemorrhage (ICH) detection on noncontrast CT (NCCT) have shown promising results but lack prospective validation. Objective: To evaluate the impact on radiologists' real-world aggregate performance for ICH detection and report turnaround times for ICH-positive examinations of a radiology department's implementation of an AI triage and notification system for ICH detection on head NCCT examinations. Methods: This prospective single-center study included adult patients who underwent head NCCT examinations from May 12, 2021 to June 30, 2021 (phase 1) or September 30, 2021 to December 4, 2021 (phase 2). Before phase 1, the radiology department implemented a commercial AI triage system for ICH detection that processed head NCCT examinations and notified radiologists of positive results through a widget with a floating pop-up display. Examinations were interpreted by neuroradiologists or emergency radiologists, who evaluated examinations without and with AI assistance in phase 1 and phase 2, respectively. A panel of radiologists conducted a review process for all examinations with discordance between the radiology report and AI and a subset of remaining examinations, to establish the reference standard. Diagnostic performance and report turnaround times were compared using Pearson chi-square test and Wilcoxon rank-sum test, respectively. Bonferroni correction was used to account for five diagnostic performance metrics (adjusted significance threshold, .01 [α=.05/5]). Results: A total of 9954 examinations from 7371 patients (mean age, 54.8±19.8 years; 3773 female, 3598 male) were included. In phases 1 and 2, 19.8% (735/3716) and 21.9% (1368/6238) of examinations, respectively, were positive for ICH (P=.01). Radiologists without versus with AI showed no significant difference in accuracy (99.5% vs 99.2%), sensitivity (98.6% vs 98.9%), PPV (99.0% vs 99.7%), or NPV (99.7% vs 99.7%) (all P>.01); specificity was higher for radiologists without than with AI (99.8% vs 99.3%, respectively, P=.004). Mean report turnaround time for ICH-positive examinations was 147.1 minutes without AI versus 149.9 minutes with AI (P=.11). Conclusion: An AI triage system for ICH detection did not improve radiologists' diagnostic performance or report turnaround times. Clinical Impact: This large prospective real-world study does not support use of AI assistance for ICH detection.

[Explainable & Safe Artificial Intelligence in Radiology]

Artificial intelligence (AI) is transforming radiology with improved diagnostic accuracy and efficiency, but prediction uncertainty remains a critical challenge. This review examines key sources of uncertainty-out-of-distribution, aleatoric, and model uncertainties-and highlights the importance of independent confidence metrics and explainable AI for safe integration. Independent confidence metrics assess the reliability of AI predictions, while explainable AI provides transparency, enhancing collaboration between AI and radiologists. The development of zero-error tolerance models, designed to minimize errors, sets new standards for safety. Addressing these challenges will enable AI to become a trusted partner in radiology, advancing care standards and patient outcomes.

Potential strength and weakness of artificial intelligence integration in emergency radiology: a review of diagnostic utilizations and applications in patient care optimization

Artificial intelligence (AI) and its recent increasing healthcare integration has created both new opportunities and challenges in the practice of radiology and medical imaging. Recent advancements in AI technology have allowed for more workplace efficiency, higher diagnostic accuracy, and overall improvements in patient care. Limitations of AI such as data imbalances, the unclear nature of AI algorithms, and the challenges in detecting certain diseases make it difficult for its widespread adoption. This review article presents cases involving the use of AI models to diagnose intracranial hemorrhage, spinal fractures, and rib fractures, while discussing how certain factors like, type, location, size, presence of artifacts, calcification, and post-surgical changes, affect AI model performance and accuracy. While the use of artificial intelligence has the potential to improve the practice of emergency radiology, it is important to address its limitations to maximize its advantages while ensuring the safety of patients overall.

Construction and Validation of a General Medical Image Dataset for Pretraining

In the field of deep learning for medical image analysis, training models from scratch are often used and sometimes, transfer learning from pretrained parameters on ImageNet models is also adopted. However, there is no universally accepted medical image dataset specifically designed for pretraining models currently. The purpose of this study is to construct such a general dataset and validate its effectiveness on downstream medical imaging tasks, including classification and segmentation. In this work, we first build a medical image dataset by collecting several public medical image datasets (CPMID). And then, some pretrained models used for transfer learning are obtained based on CPMID. Various-complexity Resnet and the Vision Transformer network are used as the backbone architectures. In the tasks of classification and segmentation on three other datasets, we compared the experimental results of training from scratch, from the pretrained parameters on ImageNet, and from the pretrained parameters on CPMID. Accuracy, the area under the receiver operating characteristic curve, and class activation map are used as metrics for classification performance. Intersection over Union as the metric is for segmentation evaluation. Utilizing the pretrained parameters on the constructed dataset CPMID, we achieved the best classification accuracy, weighted accuracy, and ROC-AUC values on three validation datasets. Notably, the average classification accuracy outperformed ImageNet-based results by 4.30%, 8.86%, and 3.85% respectively. Furthermore, we achieved the optimal balanced outcome of performance and efficiency in both classification and segmentation tasks. The pretrained parameters on the proposed dataset CPMID are very effective for common tasks in medical image analysis such as classification and segmentation.

Raman spectroscopic deep learning with signal aggregated representations for enhanced cell phenotype and signature identification

Feature representation is critical for data learning, particularly in learning spectroscopic data. Machine learning (ML) and deep learning (DL) models learn Raman spectra for rapid, nondestructive, and label-free cell phenotype identification, which facilitate diagnostic, therapeutic, forensic, and microbiological applications. But these are challenged by high-dimensional, unordered, and low-sample spectroscopic data. Here, we introduced novel 2D image-like dual signal and component aggregated representations by restructuring Raman spectra and principal components, which enables spectroscopic DL for enhanced cell phenotype and signature identification. New ConvNet models DSCARNets significantly outperformed the state-of-the-art (SOTA) ML and DL models on six benchmark datasets, mostly with >2% improvement over the SOTA performance of 85-97% accuracies. DSCARNets also performed well on four additional datasets against SOTA models of extremely high performances (>98%) and two datasets without a published supervised phenotype classification model. Explainable DSCARNets identified Raman signatures consistent with experimental indications.

Usefulness of Artificial Intelligence in Traumatic Brain Injury: A Bibliometric Analysis and Mini-review

BACKGROUND

Traumatic brain injury (TBI) has become a major source of disability worldwide, increasing the interest in algorithms that use artificial intelligence (AI) to optimize the interpretation of imaging studies, prognosis estimation, and critical care issues. In this study we present a bibliometric analysis and mini-review on the main uses that have been developed for TBI in AI.

METHODS

The results informing this review come from a Scopus database search as of April 15, 2023. The bibliometric analysis was carried out via the mapping bibliographic metrics method. Knowledge mapping was made in the VOSviewer software (V1.6.18), analyzing the "link strength" of networks based on co-occurrence of key words, countries co-authorship, and co-cited authors. In the mini-review section, we highlight the main findings and contributions of the studies.

RESULTS

A total of 495 scientific publications were identified from 2000 to 2023, with 9262 citations published since 2013. Among the 160 journals identified, The Journal of Neurotrauma, Frontiers in Neurology, and PLOS ONE were those with the greatest number of publications. The most frequently co-occurring key words were: "machine learning", "deep learning", "magnetic resonance imaging", and "intracranial pressure". The United States accounted for more collaborations than any other country, followed by United Kingdom and China. Four co-citation author clusters were found, and the top 20 papers were divided into reviews and original articles.

CONCLUSIONS

AI has become a relevant research field in TBI during the last 20 years, demonstrating great potential in imaging, but a more modest performance for prognostic estimation and neuromonitoring.

Gamified Crowdsourcing as a Novel Approach to Lung Ultrasound Data Set Labeling: Prospective Analysis

BACKGROUND

Machine learning (ML) models can yield faster and more accurate medical diagnoses; however, developing ML models is limited by a lack of high-quality labeled training data. Crowdsourced labeling is a potential solution but can be constrained by concerns about label quality.

OBJECTIVE

This study aims to examine whether a gamified crowdsourcing platform with continuous performance assessment, user feedback, and performance-based incentives could produce expert-quality labels on medical imaging data.

METHODS

In this diagnostic comparison study, 2384 lung ultrasound clips were retrospectively collected from 203 emergency department patients. A total of 6 lung ultrasound experts classified 393 of these clips as having no B-lines, one or more discrete B-lines, or confluent B-lines to create 2 sets of reference standard data sets (195 training clips and 198 test clips). Sets were respectively used to (1) train users on a gamified crowdsourcing platform and (2) compare the concordance of the resulting crowd labels to the concordance of individual experts to reference standards. Crowd opinions were sourced from DiagnosUs (Centaur Labs) iOS app users over 8 days, filtered based on past performance, aggregated using majority rule, and analyzed for label concordance compared with a hold-out test set of expert-labeled clips. The primary outcome was comparing the labeling concordance of collated crowd opinions to trained experts in classifying B-lines on lung ultrasound clips.

RESULTS

Our clinical data set included patients with a mean age of 60.0 (SD 19.0) years; 105 (51.7%) patients were female and 114 (56.1%) patients were White. Over the 195 training clips, the expert-consensus label distribution was 114 (58%) no B-lines, 56 (29%) discrete B-lines, and 25 (13%) confluent B-lines. Over the 198 test clips, expert-consensus label distribution was 138 (70%) no B-lines, 36 (18%) discrete B-lines, and 24 (12%) confluent B-lines. In total, 99,238 opinions were collected from 426 unique users. On a test set of 198 clips, the mean labeling concordance of individual experts relative to the reference standard was 85.0% (SE 2.0), compared with 87.9% crowdsourced label concordance (P=.15). When individual experts' opinions were compared with reference standard labels created by majority vote excluding their own opinion, crowd concordance was higher than the mean concordance of individual experts to reference standards (87.4% vs 80.8%, SE 1.6 for expert concordance; P<.001). Clips with discrete B-lines had the most disagreement from both the crowd consensus and individual experts with the expert consensus. Using randomly sampled subsets of crowd opinions, 7 quality-filtered opinions were sufficient to achieve near the maximum crowd concordance.

CONCLUSIONS

Crowdsourced labels for B-line classification on lung ultrasound clips via a gamified approach achieved expert-level accuracy. This suggests a strategic role for gamified crowdsourcing in efficiently generating labeled image data sets for training ML systems.

Deep learning-assisted detection and segmentation of intracranial hemorrhage in noncontrast computed tomography scans of acute stroke patients: a systematic review and meta-analysis

BACKGROUND

Deep learning (DL)-assisted detection and segmentation of intracranial hemorrhage stroke in noncontrast computed tomography (NCCT) scans are well-established, but evidence on this topic is lacking.

MATERIALS AND METHODS

PubMed and Embase databases were searched from their inception to November 2023 to identify related studies. The primary outcomes included sensitivity, specificity, and the Dice Similarity Coefficient (DSC); while the secondary outcomes were positive predictive value (PPV), negative predictive value (NPV), precision, area under the receiver operating characteristic curve (AUROC), processing time, and volume of bleeding. Random-effect model and bivariate model were used to pooled independent effect size and diagnostic meta-analysis data, respectively.

RESULTS

A total of 36 original studies were included in this meta-analysis. Pooled results indicated that DL technologies have a comparable performance in intracranial hemorrhage detection and segmentation with high values of sensitivity (0.89, 95% CI: 0.88-0.90), specificity (0.91, 95% CI: 0.89-0.93), AUROC (0.94, 95% CI: 0.93-0.95), PPV (0.92, 95% CI: 0.91-0.93), NPV (0.94, 95% CI: 0.91-0.96), precision (0.83, 95% CI: 0.77-0.90), DSC (0.84, 95% CI: 0.82-0.87). There is no significant difference between manual labeling and DL technologies in hemorrhage quantification (MD 0.08, 95% CI: -5.45-5.60, P =0.98), but the latter takes less process time than manual labeling (WMD 2.26, 95% CI: 1.96-2.56, P =0.001).

CONCLUSION

This systematic review has identified a range of DL algorithms that the performance was comparable to experienced clinicians in hemorrhage lesions identification, segmentation, and quantification but with greater efficiency and reduced cost. It is highly emphasized that multicenter randomized controlled clinical trials will be needed to validate the performance of these tools in the future, paving the way for fast and efficient decision-making during clinical procedure in patients with acute hemorrhagic stroke.

Semi-supervised Learning for Generalizable Intracranial Hemorrhage Detection and Segmentation

Purpose To develop and evaluate a semi-supervised learning model for intracranial hemorrhage detection and segmentation on an out-of-distribution head CT evaluation set. Materials and Methods This retrospective study used semi-supervised learning to bootstrap performance. An initial "teacher" deep learning model was trained on 457 pixel-labeled head CT scans collected from one U.S. institution from 2010 to 2017 and used to generate pseudo labels on a separate unlabeled corpus of 25 000 examinations from the Radiological Society of North America and American Society of Neuroradiology. A second "student" model was trained on this combined pixel- and pseudo-labeled dataset. Hyperparameter tuning was performed on a validation set of 93 scans. Testing for both classification (n = 481 examinations) and segmentation (n = 23 examinations, or 529 images) was performed on CQ500, a dataset of 481 scans performed in India, to evaluate out-of-distribution generalizability. The semi-supervised model was compared with a baseline model trained on only labeled data using area under the receiver operating characteristic curve, Dice similarity coefficient, and average precision metrics. Results The semi-supervised model achieved a statistically significant higher examination area under the receiver operating characteristic curve on CQ500 compared with the baseline (0.939 [95% CI: 0.938, 0.940] vs 0.907 [95% CI: 0.906, 0.908]; P = .009). It also achieved a higher Dice similarity coefficient (0.829 [95% CI: 0.825, 0.833] vs 0.809 [95% CI: 0.803, 0.812]; P = .012) and pixel average precision (0.848 [95% CI: 0.843, 0.853]) vs 0.828 [95% CI: 0.817, 0.828]) compared with the baseline. Conclusion The addition of unlabeled data in a semi-supervised learning framework demonstrates stronger generalizability potential for intracranial hemorrhage detection and segmentation compared with a supervised baseline. Keywords: Semi-supervised Learning, Traumatic Brain Injury, CT, Machine Learning Supplemental material is available for this article. Published under a CC BY 4.0 license. See also the commentary by Swimburne in this issue.

Understanding Bias in Artificial Intelligence: A Practice Perspective

In the fall of 2021, several experts in this space delivered a Webinar hosted by the American Society of Neuroradiology (ASNR) Diversity and Inclusion Committee, focused on expanding the understanding of bias in artificial intelligence, with a health equity lens, and provided key concepts for neuroradiologists to approach the evaluation of these tools. In this perspective, we distill key parts of this discussion, including understanding why this topic is important to neuroradiologists and lending insight on how neuroradiologists can develop a framework to assess health equity-related bias in artificial intelligence tools. In addition, we provide examples of clinical workflow implementation of these tools so that we can begin to see how artificial intelligence tools will impact discourse on equitable radiologic care. As continuous learners, we must be engaged in new and rapidly evolving technologies that emerge in our field. The Diversity and Inclusion Committee of the ASNR has addressed this subject matter through its programming content revolving around health equity in neuroradiologic advances.

Glaucoma detection model by exploiting multi-region and multi-scan-pattern OCT images with dynamical region score

Currently, deep learning-based methods have achieved success in glaucoma detection. However, most models focus on OCT images captured by a single scan pattern within a given region, holding the high risk of the omission of valuable features in the remaining regions or scan patterns. Therefore, we proposed a multi-region and multi-scan-pattern fusion model to address this issue. Our proposed model exploits comprehensive OCT images from three fundus anatomical regions (macular, middle, and optic nerve head regions) being captured by four scan patterns (radial, volume, single-line, and circular scan patterns). Moreover, to enhance the efficacy of integrating features across various scan patterns within a region and multiple regional features, we employed an attention multi-scan fusion module and an attention multi-region fusion module that auto-assign contribution to distinct scan-pattern features and region features adapting to characters of different samples, respectively. To alleviate the absence of available datasets, we have collected a specific dataset (MRMSG-OCT) comprising OCT images captured by four scan patterns from three regions. The experimental results and visualized feature maps both demonstrate that our proposed model achieves superior performance against the single scan-pattern models and single region-based models. Moreover, compared with the average fusion strategy, our proposed fusion modules yield superior performance, particularly reversing the performance degradation observed in some models relying on fixed weights, validating the efficacy of the proposed dynamic region scores adapted to different samples. Moreover, the derived region contribution scores enhance the interpretability of the model and offer an overview of the model's decision-making process, assisting ophthalmologists in prioritizing regions with heightened scores and increasing efficiency in clinical practice.

Uncertainty-aware deep-learning model for prediction of supratentorial hematoma expansion from admission non-contrast head computed tomography scan

Hematoma expansion (HE) is a modifiable risk factor and a potential treatment target in patients with intracerebral hemorrhage (ICH). We aimed to train and validate deep-learning models for high-confidence prediction of supratentorial ICH expansion, based on admission non-contrast head Computed Tomography (CT). Applying Monte Carlo dropout and entropy of deep-learning model predictions, we estimated the model uncertainty and identified patients at high risk of HE with high confidence. Using the receiver operating characteristics area under the curve (AUC), we compared the deep-learning model prediction performance with multivariable models based on visual markers of HE determined by expert reviewers. We randomly split a multicentric dataset of patients (4-to-1) into training/cross-validation (n = 634) versus test (n = 159) cohorts. We trained and tested separate models for prediction of ≥6 mL and ≥3 mL ICH expansion. The deep-learning models achieved an AUC = 0.81 for high-confidence prediction of HE≥6 mL and AUC = 0.80 for prediction of HE≥3 mL, which were higher than visual maker models AUC = 0.69 for HE≥6 mL (p = 0.036) and AUC = 0.68 for HE≥3 mL (p = 0.043). Our results show that fully automated deep-learning models can identify patients at risk of supratentorial ICH expansion based on admission non-contrast head CT, with high confidence, and more accurately than benchmark visual markers.

Examination-Level Supervision for Deep Learning-based Intracranial Hemorrhage Detection on Head CT Scans

Purpose To compare the effectiveness of weak supervision (ie, with examination-level labels only) and strong supervision (ie, with image-level labels) in training deep learning models for detection of intracranial hemorrhage (ICH) on head CT scans. Materials and Methods In this retrospective study, an attention-based convolutional neural network was trained with either local (ie, image level) or global (ie, examination level) binary labels on the Radiological Society of North America (RSNA) 2019 Brain CT Hemorrhage Challenge dataset of 21 736 examinations (8876 [40.8%] ICH) and 752 422 images (107 784 [14.3%] ICH). The CQ500 (436 examinations; 212 [48.6%] ICH) and CT-ICH (75 examinations; 36 [48.0%] ICH) datasets were employed for external testing. Performance in detecting ICH was compared between weak (examination-level labels) and strong (image-level labels) learners as a function of the number of labels available during training. Results On examination-level binary classification, strong and weak learners did not have different area under the receiver operating characteristic curve values on the internal validation split (0.96 vs 0.96; P = .64) and the CQ500 dataset (0.90 vs 0.92; P = .15). Weak learners outperformed strong ones on the CT-ICH dataset (0.95 vs 0.92; P = .03). Weak learners had better section-level ICH detection performance when more than 10 000 labels were available for training (average f1 = 0.73 vs 0.65; P < .001). Weakly supervised models trained on the entire RSNA dataset required 35 times fewer labels than equivalent strong learners. Conclusion Strongly supervised models did not achieve better performance than weakly supervised ones, which could reduce radiologist labor requirements for prospective dataset curation. Keywords: CT, Head/Neck, Brain/Brain Stem, Hemorrhage Supplemental material is available for this article. © RSNA, 2023 See also commentary by Wahid and Fuentes in this issue.

Current status of quality control in screening esophagogastroduodenoscopy and the emerging role of artificial intelligence

Esophagogastroduodenoscopy (EGD) screening is being implemented in countries with a high incidence of upper gastrointestinal (UGI) cancer. High-quality EGD screening ensures the yield of early diagnosis and prevents suffering from advanced UGI cancer and minimal operational-related discomfort. However, performance varied dramatically among endoscopists, and quality control for EGD screening remains suboptimal. Guidelines have recommended potential measures for endoscopy quality improvement and research has been conducted for evidence. Moreover, artificial intelligence offers a promising solution for computer-aided diagnosis and quality control during EGD examinations. In this review, we summarized the key points for quality assurance in EGD screening based on current guidelines and evidence. We also outline the latest evidence, limitations, and future prospects of the emerging role of artificial intelligence in EGD quality control, aiming to provide a foundation for improving the quality of EGD screening.

Application of Artificial Intelligence at All Stages of Bone Tissue Engineering

The development of artificial intelligence (AI) has revolutionized medical care in recent years and plays a vital role in a number of areas, such as diagnostics and forecasting. In this review, we discuss the most promising areas of AI application to the field of bone tissue engineering and prosthetics, which can drastically benefit from AI-assisted optimization and patient personalization of implants and scaffolds in ways ranging from visualization and real-time monitoring to the implantation cases prediction, thereby leveraging the compromise between specific architecture decisions, material choice, and synthesis procedure. With the emphasized crucial role of accuracy and robustness of developed AI algorithms, especially in bone tissue engineering, it was shown that rigorous validation and testing, demanding large datasets and extensive clinical trials, are essential, and we discuss how through developing multidisciplinary cooperation among biology, chemistry with materials science, and AI, these challenges can be addressed.

Diagnostic test accuracy of machine learning algorithms for the detection intracranial hemorrhage: a systematic review and meta-analysis study

BACKGROUND

This systematic review and meta-analysis were conducted to objectively evaluate the evidence of machine learning (ML) in the patient diagnosis of Intracranial Hemorrhage (ICH) on computed tomography (CT) scans.

METHODS

Until May 2023, systematic searches were conducted in ISI Web of Science, PubMed, Scopus, Cochrane Library, IEEE Xplore Digital Library, CINAHL, Science Direct, PROSPERO, and EMBASE for studies that evaluated the diagnostic precision of ML model-assisted ICH detection. Patients with and without ICH as the target condition who were receiving CT-Scan were eligible for the research, which used ML algorithms based on radiologists' reports as the gold reference standard. For meta-analysis, pooled sensitivities, specificities, and a summary receiver operating characteristics curve (SROC) were used.

RESULTS

At last, after screening the title, abstract, and full paper, twenty-six retrospective and three prospective, and two retrospective/prospective studies were included. The overall (Diagnostic Test Accuracy) DTA of retrospective studies with a pooled sensitivity was 0.917 (95% CI 0.88-0.943, I2 = 99%). The pooled specificity was 0.945 (95% CI 0.918-0.964, I2 = 100%). The pooled diagnostic odds ratio (DOR) was 219.47 (95% CI 104.78-459.66, I2 = 100%). These results were significant for the specificity of the different network architecture models (p-value = 0.0289). However, the results for sensitivity (p-value = 0.6417) and DOR (p-value = 0.2187) were not significant. The ResNet algorithm has higher pooled specificity than other algorithms with 0.935 (95% CI 0.854-0.973, I2 = 93%).

CONCLUSION

This meta-analysis on DTA of ML algorithms for detecting ICH by assessing non-contrast CT-Scans shows the ML has an acceptable performance in diagnosing ICH. Using ResNet in ICH detection remains promising prediction was improved via training in an Architecture Learning Network (ALN).

Insights on the Current State and Future Outlook of AI in Health Care: Expert Interview Study

BACKGROUND

Artificial intelligence (AI) is often promoted as a potential solution for many challenges health care systems face worldwide. However, its implementation in clinical practice lags behind its technological development.

OBJECTIVE

This study aims to gain insights into the current state and prospects of AI technology from the stakeholders most directly involved in its adoption in the health care sector whose perspectives have received limited attention in research to date.

METHODS

For this purpose, the perspectives of AI researchers and health care IT professionals in North America and Western Europe were collected and compared for profession-specific and regional differences. In this preregistered, mixed methods, cross-sectional study, 23 experts were interviewed using a semistructured guide. Data from the interviews were analyzed using deductive and inductive qualitative methods for the thematic analysis along with topic modeling to identify latent topics.

RESULTS

Through our thematic analysis, four major categories emerged: (1) the current state of AI systems in health care, (2) the criteria and requirements for implementing AI systems in health care, (3) the challenges in implementing AI systems in health care, and (4) the prospects of the technology. Experts discussed the capabilities and limitations of current AI systems in health care in addition to their prevalence and regional differences. Several criteria and requirements deemed necessary for the successful implementation of AI systems were identified, including the technology's performance and security, smooth system integration and human-AI interaction, costs, stakeholder involvement, and employee training. However, regulatory, logistical, and technical issues were identified as the most critical barriers to an effective technology implementation process. In the future, our experts predicted both various threats and many opportunities related to AI technology in the health care sector.

CONCLUSIONS

Our work provides new insights into the current state, criteria, challenges, and outlook for implementing AI technology in health care from the perspective of AI researchers and IT professionals in North America and Western Europe. For the full potential of AI-enabled technologies to be exploited and for them to contribute to solving current health care challenges, critical implementation criteria must be met, and all groups involved in the process must work together.

Development and Validation of a Deep Learning-Based Histologic Diagnosis System for Diagnosing Colorectal Sessile Serrated Lesions

OBJECTIVES

The histopathologic diagnosis of colorectal sessile serrated lesions (SSLs) and hyperplastic polyps (HPs) is of low consistency among pathologists. This study aimed to develop and validate a deep learning (DL)-based logical anthropomorphic pathology diagnostic system (LA-SSLD) for the differential diagnosis of colorectal SSL and HP.

METHODS

The diagnosis framework of the LA-SSLD system was constructed according to the current guidelines and consisted of 4 DL models. Deep convolutional neural network (DCNN) 1 was the mucosal layer segmentation model, DCNN 2 was the muscularis mucosa segmentation model, DCNN 3 was the glandular lumen segmentation model, and DCNN 4 was the glandular lumen classification (aberrant or regular) model. A total of 175 HP and 127 SSL sections were collected from Renmin Hospital of Wuhan University during November 2016 to November 2022. The performance of the LA-SSLD system was compared to 11 pathologists with different qualifications through the human-machine contest.

RESULTS

The Dice scores of DCNNs 1, 2, and 3 were 93.66%, 58.38%, and 74.04%, respectively. The accuracy of DCNN 4 was 92.72%. In the human-machine contest, the accuracy, sensitivity, and specificity of the LA-SSLD system were 85.71%, 86.36%, and 85.00%, respectively. In comparison with experts (pathologist D: accuracy 83.33%, sensitivity 90.91%, specificity 75.00%; pathologist E: accuracy 85.71%, sensitivity 90.91%, specificity 80.00%), LA-SSLD achieved expert-level accuracy and outperformed all the senior and junior pathologists.

CONCLUSIONS

This study proposed a logical anthropomorphic diagnostic system for the differential diagnosis of colorectal SSL and HP. The diagnostic performance of the system is comparable to that of experts and has the potential to become a powerful diagnostic tool for SSL in the future. It is worth mentioning that a logical anthropomorphic system can achieve expert-level accuracy with fewer samples, providing potential ideas for the development of other artificial intelligence models.

Application of deep learning models for detection of subdural hematoma: a systematic review and meta-analysis

BACKGROUND

This study aimed to investigate the application of deep learning (DL) models for the detection of subdural hematoma (SDH).

METHODS

We conducted a comprehensive search using relevant keywords. Articles extracted were original studies in which sensitivity and/or specificity were reported. Two different approaches of frequentist and Bayesian inference were applied. For quality and risk of bias assessment we used Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2).

RESULTS

We analyzed 22 articles that included 1,997,749 patients. In the first step, the frequentist method showed a pooled sensitivity of 88.8% (95% confidence interval (CI): 83.9% to 92.4%) and a specificity of 97.2% (95% CI 94.6% to 98.6%). In the second step, using Bayesian methods including 11 studies that reported sensitivity and specificity, a sensitivity rate of 86.8% (95% CI: 77.6% to 92.9%) at a specificity level of 86.9% (95% CI: 60.9% to 97.2%) was achieved. The risk of bias assessment was not remarkable using QUADAS-2.

CONCLUSION

DL models might be an appropriate tool for detecting SDHs with a reasonably high sensitivity and specificity.

Leveraging physiology and artificial intelligence to deliver advancements in health care

Artificial intelligence in health care has experienced remarkable innovation and progress in the last decade. Significant advancements can be attributed to the utilization of artificial intelligence to transform physiology data to advance health care. In this review, we explore how past work has shaped the field and defined future challenges and directions. In particular, we focus on three areas of development. First, we give an overview of artificial intelligence, with special attention to the most relevant artificial intelligence models. We then detail how physiology data have been harnessed by artificial intelligence to advance the main areas of health care: automating existing health care tasks, increasing access to care, and augmenting health care capabilities. Finally, we discuss emerging concerns surrounding the use of individual physiology data and detail an increasingly important consideration for the field, namely the challenges of deploying artificial intelligence models to achieve meaningful clinical impact.

Comparison and summary of in silico prediction tools for CYP450-mediated drug metabolism

The cytochrome P450 (CYP450) enzyme system is responsible for the metabolism of more than two-thirds of xenobiotics. This review summarizes reports of a series of in silico tools for CYP450 enzyme-drug interaction predictions, including the prediction of sites of metabolism (SOM) of a drug and the identification of inhibitor/substrates for CYP subtypes. We also evaluated four prediction tools to identify CYP inhibitors utilizing 52 of the most frequently prescribed drugs. ADMET Predictor and CYPlebrity demonstrated the best performance. We hope that this review provides guidance for choosing appropriate enzyme prediction tools from a variety of in silico platforms to meet individual needs. Such predictions are useful for medicinal chemists to prioritize their designed compounds for further drug discovery.

Hybrid morphological-convolutional neural networks for computer-aided diagnosis

Training deep Convolutional Neural Networks (CNNs) presents challenges in terms of memory requirements and computational resources, often resulting in issues such as model overfitting and lack of generalization. These challenges can only be mitigated by using an excessive number of training images. However, medical image datasets commonly suffer from data scarcity due to the complexities involved in their acquisition, preparation, and curation. To address this issue, we propose a compact and hybrid machine learning architecture based on the Morphological and Convolutional Neural Network (MCNN), followed by a Random Forest classifier. Unlike deep CNN architectures, the MCNN was specifically designed to achieve effective performance with medical image datasets limited to a few hundred samples. It incorporates various morphological operations into a single layer and uses independent neural networks to extract information from each signal channel. The final classification is obtained by utilizing a Random Forest classifier on the outputs of the last neural network layer. We compare the classification performance of our proposed method with three popular deep CNN architectures (ResNet-18, ShuffleNet-V2, and MobileNet-V2) using two training approaches: full training and transfer learning. The evaluation was conducted on two distinct medical image datasets: the ISIC dataset for melanoma classification and the ORIGA dataset for glaucoma classification. Results demonstrate that the MCNN method exhibits reliable performance in melanoma classification, achieving an AUC of 0.94 (95% CI: 0.91 to 0.97), outperforming the popular CNN architectures. For the glaucoma dataset, the MCNN achieved an AUC of 0.65 (95% CI: 0.53 to 0.74), which was similar to the performance of the popular CNN architectures. This study contributes to the understanding of mathematical morphology in shallow neural networks for medical image classification and highlights the potential of hybrid architectures in effectively learning from medical image datasets that are limited by a small number of case samples.

Explainable deep learning in plant phenotyping

The increasing human population and variable weather conditions, due to climate change, pose a threat to the world's food security. To improve global food security, we need to provide breeders with tools to develop crop cultivars that are more resilient to extreme weather conditions and provide growers with tools to more effectively manage biotic and abiotic stresses in their crops. Plant phenotyping, the measurement of a plant's structural and functional characteristics, has the potential to inform, improve and accelerate both breeders' selections and growers' management decisions. To improve the speed, reliability and scale of plant phenotyping procedures, many researchers have adopted deep learning methods to estimate phenotypic information from images of plants and crops. Despite the successful results of these image-based phenotyping studies, the representations learned by deep learning models remain difficult to interpret, understand, and explain. For this reason, deep learning models are still considered to be black boxes. Explainable AI (XAI) is a promising approach for opening the deep learning model's black box and providing plant scientists with image-based phenotypic information that is interpretable and trustworthy. Although various fields of study have adopted XAI to advance their understanding of deep learning models, it has yet to be well-studied in the context of plant phenotyping research. In this review article, we reviewed existing XAI studies in plant shoot phenotyping, as well as related domains, to help plant researchers understand the benefits of XAI and make it easier for them to integrate XAI into their future studies. An elucidation of the representations within a deep learning model can help researchers explain the model's decisions, relate the features detected by the model to the underlying plant physiology, and enhance the trustworthiness of image-based phenotypic information used in food production systems.

A translational perspective towards clinical AI fairness

Artificial intelligence (AI) has demonstrated the ability to extract insights from data, but the fairness of such data-driven insights remains a concern in high-stakes fields. Despite extensive developments, issues of AI fairness in clinical contexts have not been adequately addressed. A fair model is normally expected to perform equally across subgroups defined by sensitive variables (e.g., age, gender/sex, race/ethnicity, socio-economic status, etc.). Various fairness measurements have been developed to detect differences between subgroups as evidence of bias, and bias mitigation methods are designed to reduce the differences detected. This perspective of fairness, however, is misaligned with some key considerations in clinical contexts. The set of sensitive variables used in healthcare applications must be carefully examined for relevance and justified by clear clinical motivations. In addition, clinical AI fairness should closely investigate the ethical implications of fairness measurements (e.g., potential conflicts between group- and individual-level fairness) to select suitable and objective metrics. Generally defining AI fairness as "equality" is not necessarily reasonable in clinical settings, as differences may have clinical justifications and do not indicate biases. Instead, "equity" would be an appropriate objective of clinical AI fairness. Moreover, clinical feedback is essential to developing fair and well-performing AI models, and efforts should be made to actively involve clinicians in the process. The adaptation of AI fairness towards healthcare is not self-evident due to misalignments between technical developments and clinical considerations. Multidisciplinary collaboration between AI researchers, clinicians, and ethicists is necessary to bridge the gap and translate AI fairness into real-life benefits.

A historical perspective of biomedical explainable AI research

The black-box nature of most artificial intelligence (AI) models encourages the development of explainability methods to engender trust into the AI decision-making process. Such methods can be broadly categorized into two main types: post hoc explanations and inherently interpretable algorithms. We aimed at analyzing the possible associations between COVID-19 and the push of explainable AI (XAI) to the forefront of biomedical research. We automatically extracted from the PubMed database biomedical XAI studies related to concepts of causality or explainability and manually labeled 1,603 papers with respect to XAI categories. To compare the trends pre- and post-COVID-19, we fit a change point detection model and evaluated significant changes in publication rates. We show that the advent of COVID-19 in the beginning of 2020 could be the driving factor behind an increased focus concerning XAI, playing a crucial role in accelerating an already evolving trend. Finally, we present a discussion with future societal use and impact of XAI technologies and potential future directions for those who pursue fostering clinical trust with interpretable machine learning models.

Deep learning of 2D-Restructured gene expression representations for improved low-sample therapeutic response prediction

Clinical outcome prediction is important for stratified therapeutics. Machine learning (ML) and deep learning (DL) methods facilitate therapeutic response prediction from transcriptomic profiles of cells and clinical samples. Clinical transcriptomic DL is challenged by the low-sample sizes (34-286 subjects), high-dimensionality (up to 21,653 genes) and unordered nature of clinical transcriptomic data. The established methods rely on ML algorithms at accuracy levels of 0.6-0.8 AUC/ACC values. Low-sample DL algorithms are needed for enhanced prediction capability. Here, an unsupervised manifold-guided algorithm was employed for restructuring transcriptomic data into ordered image-like 2D-representations, followed by efficient DL of these 2D-representations with deep ConvNets. Our DL models significantly outperformed the state-of-the-art (SOTA) ML models on 82% of 17 low-sample benchmark datasets (53% with >0.05 AUC/ACC improvement). They are more robust than the SOTA models in cross-cohort prediction tasks, and in identifying robust biomarkers and response-dependent variational patterns consistent with experimental indications.

Improvement of a prediction model for heart failure survival through explainable artificial intelligence

Cardiovascular diseases and their associated disorder of heart failure (HF) are major causes of death globally, making it a priority for doctors to detect and predict their onset and medical consequences. Artificial Intelligence (AI) allows doctors to discover clinical indicators and enhance their diagnoses and treatments. Specifically, "eXplainable AI" (XAI) offers tools to improve the clinical prediction models that experience poor interpretability of their results. This work presents an explainability analysis and evaluation of two HF survival prediction models using a dataset that includes 299 patients who have experienced HF. The first model utilizes survival analysis, considering death events and time as target features, while the second model approaches the problem as a classification task to predict death. The model employs an optimization data workflow pipeline capable of selecting the best machine learning algorithm as well as the optimal collection of features. Moreover, different post hoc techniques have been used for the explainability analysis of the model. The main contribution of this paper is an explainability-driven approach to select the best HF survival prediction model balancing prediction performance and explainability. Therefore, the most balanced explainable prediction models are Survival Gradient Boosting model for the survival analysis and Random Forest for the classification approach with a c-index of 0.714 and balanced accuracy of 0.74 (std 0.03) respectively. The selection of features by the SCI-XAI in the two models is similar where "serum_creatinine", "ejection_fraction", and "sex" are selected in both approaches, with the addition of "diabetes" for the survival analysis model. Moreover, the application of post hoc XAI techniques also confirm common findings from both approaches by placing the "serum_creatinine" as the most relevant feature for the predicted outcome, followed by "ejection_fraction". The explainable prediction models for HF survival presented in this paper would improve the further adoption of clinical prediction models by providing doctors with insights to better understand the reasoning behind usually "black-box" AI clinical solutions and make more reasonable and data-driven decisions.

Deep learning referral suggestion and tumour discrimination using explainable artificial intelligence applied to multiparametric MRI

OBJECTIVES

An appropriate and fast clinical referral suggestion is important for intra-axial mass-like lesions (IMLLs) in the emergency setting. We aimed to apply an interpretable deep learning (DL) system to multiparametric MRI to obtain clinical referral suggestion for IMLLs, and to validate it in the setting of nontraumatic emergency neuroradiology.

METHODS

A DL system was developed in 747 patients with IMLLs ranging 30 diseases who underwent pre- and post-contrast T1-weighted (T1CE), FLAIR, and diffusion-weighted imaging (DWI). A DL system that segments IMLLs, classifies tumourous conditions, and suggests clinical referral among surgery, systematic work-up, medical treatment, and conservative treatment, was developed. The system was validated in an independent cohort of 130 emergency patients, and performance in referral suggestion and tumour discrimination was compared with that of radiologists using receiver operating characteristics curve, precision-recall curve analysis, and confusion matrices. Multiparametric interpretable visualisation of high-relevance regions from layer-wise relevance propagation overlaid on contrast-enhanced T1WI and DWI was analysed.

RESULTS

The DL system provided correct referral suggestions in 94 of 130 patients (72.3%) and performed comparably to radiologists (accuracy 72.6%, McNemar test; p = .942). For distinguishing tumours from non-tumourous conditions, the DL system (AUC, 0.90 and AUPRC, 0.94) performed similarly to human readers (AUC, 0.81~0.92, and AUPRC, 0.88~0.95). Solid portions of tumours showed a high overlap of relevance, but non-tumours did not (Dice coefficient 0.77 vs. 0.33, p < .001), demonstrating the DL's decision.

CONCLUSIONS

Our DL system could appropriately triage patients using multiparametric MRI and provide interpretability through multiparametric heatmaps, and may thereby aid neuroradiologic diagnoses in emergency settings.

CLINICAL RELEVANCE STATEMENT

Our AI triages patients with raw MRI images to clinical referral pathways in brain intra-axial mass-like lesions. We demonstrate that the decision is based on the relative relevance between contrast-enhanced T1-weighted and diffusion-weighted images, providing explainability across multiparametric MRI data.

KEY POINTS

• A deep learning (DL) system using multiparametric MRI suggested clinical referral to patients with intra-axial mass-like lesions (IMLLs) similar to radiologists (accuracy 72.3% vs. 72.6%). • In the differentiation of tumourous and non-tumourous conditions, the DL system (AUC, 0.90) performed similar with radiologists (AUC, 0.81-0.92). • The DL's decision basis for differentiating tumours from non-tumours can be quantified using multiparametric heatmaps obtained via the layer-wise relevance propagation method.

An Explainable Vision Transformer Model Based White Blood Cells Classification and Localization

White blood cells (WBCs) are crucial components of the immune system that play a vital role in defending the body against infections and diseases. The identification of WBCs subtypes is useful in the detection of various diseases, such as infections, leukemia, and other hematological malignancies. The manual screening of blood films is time-consuming and subjective, leading to inconsistencies and errors. Convolutional neural networks (CNN)-based models can automate such classification processes, but are incapable of capturing long-range dependencies and global context. This paper proposes an explainable Vision Transformer (ViT) model for automatic WBCs detection from blood films. The proposed model uses a self-attention mechanism to extract features from input images. Our proposed model was trained and validated on a public dataset of 16,633 samples containing five different types of WBCs. As a result of experiments on the classification of five different types of WBCs, our model achieved an accuracy of 99.40%. Moreover, the model's examination of misclassified test samples revealed a correlation between incorrect predictions and the presence or absence of granules in the cell samples. To validate this observation, we divided the dataset into two classes, Granulocytes and Agranulocytes, and conducted a secondary training process. The resulting ViT model, trained for binary classification, achieved impressive performance metrics during the test phase, including an accuracy of 99.70%, recall of 99.54%, precision of 99.32%, and F-1 score of 99.43%. To ensure the reliability of the ViT model's, we employed the Score-CAM algorithm to visualize the pixel areas on which the model focuses during its predictions. Our proposed method is suitable for clinical use due to its explainable structure as well as its superior performance compared to similar studies in the literature. The classification and localization of WBCs with this model can facilitate the detection and reporting process for the pathologist.

Intracranial hemorrhage detection in 3D computed tomography images using a bi-directional long short-term memory network-based modified genetic algorithm

Introduction

Intracranial hemorrhage detection in 3D Computed Tomography (CT) brain images has gained more attention in the research community. The major issue to deal with the 3D CT brain images is scarce and hard to obtain the labelled data with better recognition results.

Methods

To overcome the aforementioned problem, a new model has been implemented in this research manuscript. After acquiring the images from the Radiological Society of North America (RSNA) 2019 database, the region of interest (RoI) was segmented by employing Otsu's thresholding method. Then, feature extraction was performed utilizing Tamura features: directionality, contrast, coarseness, and Gradient Local Ternary Pattern (GLTP) descriptors to extract vectors from the segmented RoI regions. The extracted vectors were dimensionally reduced by proposing a modified genetic algorithm, where the infinite feature selection technique was incorporated with the conventional genetic algorithm to further reduce the redundancy within the regularized vectors. The selected optimal vectors were finally fed to the Bi-directional Long Short Term Memory (Bi-LSTM) network to classify intracranial hemorrhage sub-types, such as subdural, intraparenchymal, subarachnoid, epidural, and intraventricular.

Results

The experimental investigation demonstrated that the Bi-LSTM based modified genetic algorithm obtained 99.40% sensitivity, 99.80% accuracy, and 99.48% specificity, which are higher compared to the existing machine learning models: Naïve Bayes, Random Forest, Support Vector Machine (SVM), Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM) network.

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.

Automated patellar height assessment on high-resolution radiographs with a novel deep learning-based approach

BACKGROUND

Artificial intelligence and deep learning have shown promising results in medical imaging and interpreting radiographs. Moreover, medical community shows a gaining interest in automating routine diagnostics issues and orthopedic measurements.

AIM

To verify the accuracy of automated patellar height assessment using deep learning-based bone segmentation and detection approach on high resolution radiographs.

METHODS

218 Lateral knee radiographs were included in the analysis. 82 radiographs were utilized for training and 10 other radiographs for validation of a U-Net neural network to achieve required Dice score. 92 other radiographs were used for automatic (U-Net) and manual measurements of the patellar height, quantified by Caton-Deschamps (CD) and Blackburne-Peel (BP) indexes. The detection of required bones regions on high-resolution images was done using a You Only Look Once (YOLO) neural network. The agreement between manual and automatic measurements was calculated using the interclass correlation coefficient (ICC) and the standard error for single measurement (SEM). To check U-Net's generalization the segmentation accuracy on the test set was also calculated.

RESULTS

Proximal tibia and patella was segmented with accuracy 95.9% (Dice score) by U-Net neural network on lateral knee subimages automatically detected by the YOLO network (mean Average Precision mAP greater than 0.96). The mean values of CD and BP indexes calculated by orthopedic surgeons (R#1 and R#2) was 0.93 (± 0.19) and 0.89 (± 0.19) for CD and 0.80 (± 0.17) and 0.78 (± 0.17) for BP. Automatic measurements performed by our algorithm for CD and BP indexes were 0.92 (± 0.21) and 0.75 (± 0.19), respectively. Excellent agreement between the orthopedic surgeons’ measurements and results of the algorithm has been achieved (ICC > 0.75, SEM < 0.014).

CONCLUSION

Automatic patellar height assessment can be achieved on high-resolution radiographs with the required accuracy. Determining patellar end-points and the joint line-fitting to the proximal tibia joint surface allows for accurate CD and BP index calculations. The obtained results indicate that this approach can be valuable tool in a medical practice.

Explainable deep learning-based clinical decision support engine for MRI-based automated diagnosis of temporomandibular joint anterior disk displacement

BACKGROUND AND OBJECTIVE

MRI is considered the gold standard for diagnosing anterior disc displacement (ADD), the most common temporomandibular joint (TMJ) disorder. However, even highly trained clinicians find it difficult to integrate the dynamic nature of MRI with the complicated anatomical features of the TMJ. As the first validated study for MRI-based automatic TMJ ADD diagnosis, we propose a clinical decision support engine that diagnoses TMJ ADD using MR images and provides heat maps as the visualized rationale of diagnostic predictions using explainable artificial intelligence.

METHODS

The engine builds on two deep learning models. The first deep learning model detects a region of interest (ROI) containing three TMJ components (i.e., temporal bone, disc, and condyle) in the entire sagittal MR image. The second deep learning model classifies TMJ ADD into three classes (i.e., normal, ADD without reduction, and ADD with reduction) within the detected ROI. In this retrospective study, the models were developed and tested on the dataset acquired between April 2005 to April 2020. The additional independent dataset acquired at a different hospital between January 2016 to February 2019 was used for the external test of the classification model. Detection performance was assessed by mean average precision (mAP). Classification performance was assessed by the area under the receiver operating characteristic (AUROC), sensitivity, specificity, and Youden's index. 95% confidence intervals were calculated via non-parametric bootstrap to assess the statistical significance of model performances.

RESULTS

The ROI detection model achieved mAP of 0.819 at 0.75 intersection over union (IoU) thresholds in the internal test. In internal and external tests, the ADD classification model achieved AUROC values of 0.985 and 0.960, sensitivities of 0.950 and 0.926, and specificities of 0.919 and 0.892, respectively.

CONCLUSIONS

The proposed explainable deep learning-based engine provides clinicians with the predictive result and its visualized rationale. The clinicians can make the final diagnosis by integrating primary diagnostic prediction obtained from the proposed engine with the patient's clinical examination findings.

Polycystic Ovary Syndrome Detection Machine Learning Model Based on Optimized Feature Selection and Explainable Artificial Intelligence

Polycystic ovary syndrome (PCOS) has been classified as a severe health problem common among women globally. Early detection and treatment of PCOS reduce the possibility of long-term complications, such as increasing the chances of developing type 2 diabetes and gestational diabetes. Therefore, effective and early PCOS diagnosis will help the healthcare systems to reduce the disease's problems and complications. Machine learning (ML) and ensemble learning have recently shown promising results in medical diagnostics. The main goal of our research is to provide model explanations to ensure efficiency, effectiveness, and trust in the developed model through local and global explanations. Feature selection methods with different types of ML models (logistic regression (LR), random forest (RF), decision tree (DT), naive Bayes (NB), support vector machine (SVM), k-nearest neighbor (KNN), xgboost, and Adaboost algorithm to get optimal feature selection and best model. Stacking ML models that combine the best base ML models with meta-learner are proposed to improve performance. Bayesian optimization is used to optimize ML models. Combining SMOTE (Synthetic Minority Oversampling Techniques) and ENN (Edited Nearest Neighbour) solves the class imbalance. The experimental results were made using a benchmark PCOS dataset with two ratios splitting 70:30 and 80:20. The result showed that the Stacking ML with REF feature selection recorded the highest accuracy at 100 compared to other models.

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.

Evaluation of techniques to improve a deep learning algorithm for the automatic detection of intracranial haemorrhage on CT head imaging

BACKGROUND

Deep learning (DL) algorithms are playing an increasing role in automatic medical image analysis.

PURPOSE

To evaluate the performance of a DL model for the automatic detection of intracranial haemorrhage and its subtypes on non-contrast CT (NCCT) head studies and to compare the effects of various preprocessing and model design implementations.

METHODS

The DL algorithm was trained and externally validated on open-source, multi-centre retrospective data containing radiologist-annotated NCCT head studies. The training dataset was sourced from four research institutions across Canada, the USA and Brazil. The test dataset was sourced from a research centre in India. A convolutional neural network (CNN) was used, with its performance compared against similar models with additional implementations: (1) a recurrent neural network (RNN) attached to the CNN, (2) preprocessed CT image-windowed inputs and (3) preprocessed CT image-concatenated inputs. The area under the receiver operating characteristic curve (AUC-ROC) and microaveraged precision (mAP) score were used to evaluate and compare model performances.

RESULTS

The training and test datasets contained 21,744 and 491 NCCT head studies, respectively, with 8,882 (40.8%) and 205 (41.8%) positive for intracranial haemorrhage. Implementation of preprocessing techniques and the CNN-RNN framework increased mAP from 0.77 to 0.93 and increased AUC-ROC [95% confidence intervals] from 0.854 [0.816-0.889] to 0.966 [0.951-0.980] (p-value = 3.91 × 10^-12).

CONCLUSIONS

The deep learning model accurately detected intracranial haemorrhage and improved in performance following specific implementation techniques, demonstrating clinical potential as a decision support tool and an automated system to improve radiologist workflow efficiency.

KEY POINTS

• The deep learning model detected intracranial haemorrhages on computed tomography with high accuracy. • Image preprocessing, such as windowing, plays a large role in improving deep learning model performance. • Implementations which enable an analysis of interslice dependencies can improve deep learning model performance. • Visual saliency maps can facilitate explainable artificial intelligence systems. • Deep learning within a triage system may expedite earlier intracranial haemorrhage detection.

Deep learning based automatic detection algorithm for acute intracranial haemorrhage: a pivotal randomized clinical trial

Acute intracranial haemorrhage (AIH) is a potentially life-threatening emergency that requires prompt and accurate assessment and management. This study aims to develop and validate an artificial intelligence (AI) algorithm for diagnosing AIH using brain-computed tomography (CT) images. A retrospective, multi-reader, pivotal, crossover, randomised study was performed to validate the performance of an AI algorithm was trained using 104,666 slices from 3010 patients. Brain CT images (12,663 slices from 296 patients) were evaluated by nine reviewers belonging to one of the three subgroups (non-radiologist physicians, n = 3; board-certified radiologists, n = 3; and neuroradiologists, n = 3) with and without the aid of our AI algorithm. Sensitivity, specificity, and accuracy were compared between AI-unassisted and AI-assisted interpretations using the chi-square test. Brain CT interpretation with AI assistance results in significantly higher diagnostic accuracy than that without AI assistance (0.9703 vs. 0.9471, p < 0.0001, patient-wise). Among the three subgroups of reviewers, non-radiologist physicians demonstrate the greatest improvement in diagnostic accuracy for brain CT interpretation with AI assistance compared to that without AI assistance. For board-certified radiologists, the diagnostic accuracy for brain CT interpretation is significantly higher with AI assistance than without AI assistance. For neuroradiologists, although brain CT interpretation with AI assistance results in a trend for higher diagnostic accuracy compared to that without AI assistance, the difference does not reach statistical significance. For the detection of AIH, brain CT interpretation with AI assistance results in better diagnostic performance than that without AI assistance, with the most significant improvement observed for non-radiologist physicians.

An efficient edge/cloud medical system for rapid detection of level of consciousness in emergency medicine based on explainable machine learning models

Emergency medicine (EM) is one of the attractive research fields in which researchers investigate their efforts to diagnose and treat unforeseen illnesses or injuries. There are many tests and observations are involved in EM. Detection of the level of consciousness is one of these observations, which can be detected using several methods. Among these methods, the automatic estimation of the Glasgow coma scale (GCS) is studied in this paper. The GCS is a medical score used to describe a patient’s level of consciousness. This type of scoring system requires medical examination that may not be available with the shortage of the medical expert. Therefore, the automatic medical calculation for a patient’s level of consciousness is highly needed. Artificial intelligence has been deployed in several applications and appears to have a high performance regarding providing automatic solutions. The main objective of this work is to introduce the edge/cloud system to improve the efficiency of the consciousness measurement through efficient local data processing. Moreover, an efficient machine learning (ML) model to predict the level of consciousness of a certain patient based on the patient’s demographic, vital signs, and laboratory tests is proposed, as well as maintaining the explainability issue using Shapley additive explanations (SHAP) that provides natural language explanation in a form that helps the medical expert to understand the final prediction. The developed ML model is validated using vital signs and laboratory tests extracted from the MIMIC III dataset, and it achieves superior performance (mean absolute error (MAE) = 0.269, mean square error (MSE) = 0.625, R² score = 0.964). The resulting model is accurate, medically intuitive, and trustworthy.

Prior knowledge-based precise diagnosis of blend sign from head computed tomography

Introduction

Automated diagnosis of intracranial hemorrhage on head computed tomography (CT) plays a decisive role in clinical management. This paper presents a prior knowledge-based precise diagnosis of blend sign network from head CT scans.

Method

We employ the object detection task as an auxiliary task in addition to the classification task, which could incorporate the hemorrhage location as prior knowledge into the detection framework. The auxiliary task could help the model pay more attention to the regions with hemorrhage, which is beneficial for distinguishing the blend sign. Furthermore, we propose a self-knowledge distillation strategy to deal with inaccuracy annotations.

Results

In the experiment, we retrospectively collected 1749 anonymous non-contrast head CT scans from the First Affiliated Hospital of China Medical University. The dataset contains three categories: no intracranial hemorrhage (non-ICH), normal intracranial hemorrhage (normal ICH), and blend sign. The experimental results demonstrate that our method performs better than other methods.

Discussion

Our method has the potential to assist less-experienced head CT interpreters, reduce radiologists' workload, and improve efficiency in natural clinical settings.

Intracranial Hemorrhage Detection Using Parallel Deep Convolutional Models and Boosting Mechanism

Intracranial hemorrhage (ICH) can lead to death or disability, which requires immediate action from radiologists. Due to the heavy workload, less experienced staff, and the complexity of subtle hemorrhages, a more intelligent and automated system is necessary to detect ICH. In literature, many artificial-intelligence-based methods are proposed. However, they are less accurate for ICH detection and subtype classification. Therefore, in this paper, we present a new methodology to improve the detection and subtype classification of ICH based on two parallel paths and a boosting technique. The first path employs the architecture of ResNet101-V2 to extract potential features from windowed slices, whereas Inception-V4 captures significant spatial information in the second path. Afterwards, the detection and subtype classification of ICH is performed by the light gradient boosting machine (LGBM) using the outputs of ResNet101-V2 and Inception-V4. Thus, the combined solution, known as ResNet101-V2, Inception-V4, and LGBM (Res-Inc-LGBM), is trained and tested over the brain computed tomography (CT) scans of CQ500 and Radiological Society of North America (RSNA) datasets. The experimental results state that the proposed solution efficiently obtains 97.7% accuracy, 96.5% sensitivity, and 97.4% F1 score using the RSNA dataset. Moreover, the proposed Res-Inc-LGBM outperforms the standard benchmarks for the detection and subtype classification of ICH regarding the accuracy, sensitivity, and F1 score. The results prove the significance of the proposed solution for its real-time application.

Interpretable deep learning for the prognosis of long-term functional outcome post-stroke using acute diffusion weighted imaging

Advances in deep learning can be applied to acute stroke imaging to build powerful and explainable prediction models that could supersede traditionally used biomarkers. We aimed to evaluate the performance and interpretability of a deep learning model based on convolutional neural networks (CNN) in predicting long-term functional outcome with diffusion-weighted imaging (DWI) acquired at day 1 post-stroke. Ischemic stroke patients (n = 322) were included from the ASTER and INSULINFARCT trials as well as the Pitié-Salpêtrière registry. We trained a CNN to predict long-term functional outcome assessed at 3 months with the modified Rankin Scale (dichotomized as good [mRS ≤ 2] vs. poor [mRS ≥ 3]) and compared its performance to two logistic regression models using lesion volume and ASPECTS. The CNN contained an attention mechanism, which allowed to visualize the areas of the brain that drove prediction. The deep learning model yielded a significantly higher area under the curve (0.83 95%CI [0.78-0.87]) than lesion volume (0.78 [0.73-0.83]) and ASPECTS (0.77 [0.71-0.83]) (p < 0.05). Setting all classifiers to the specificity as the deep learning model (i.e., 0.87 [0.82-0.92]), the CNN yielded a significantly higher sensitivity (0.67 [0.59-0.73]) than lesion volume (0.48 [0.40-0.56]) and ASPECTS (0.50 [0.41-0.58]) (p = 0.002). The attention mechanism revealed that the network learned to naturally attend to the lesion to predict outcome.

Multi-scale, domain knowledge-guided attention + random forest: a two-stage deep learning-based multi-scale guided attention models to diagnose idiopathic pulmonary fibrosis from computed tomography images

BACKGROUND

Idiopathic pulmonary fibrosis (IPF) is a progressive, irreversible, and usually fatal lung disease of unknown reasons, generally affecting the elderly population. Early diagnosis of IPF is crucial for triaging patients' treatment planning into anti-fibrotic treatment or treatments for other causes of pulmonary fibrosis. However, current IPF diagnosis workflow is complicated and time-consuming, which involves collaborative efforts from radiologists, pathologists, and clinicians and it is largely subject to inter-observer variability.

PURPOSE

The purpose of this work is to develop a deep learning-based automated system that can diagnose subjects with IPF among subjects with interstitial lung disease (ILD) using an axial chest computed tomography (CT) scan. This work can potentially enable timely diagnosis decisions and reduce inter-observer variability.

METHODS

Our dataset contains CT scans from 349 IPF patients and 529 non-IPF ILD patients. We used 80% of the dataset for training and validation purposes and 20% as the holdout test set. We proposed a two-stage model: at stage one, we built a multi-scale, domain knowledge-guided attention model (MSGA) that encouraged the model to focus on specific areas of interest to enhance model explainability, including both high- and medium-resolution attentions; at stage two, we collected the output from MSGA and constructed a random forest (RF) classifier for patient-level diagnosis, to further boost model accuracy. RF classifier is utilized as a final decision stage since it is interpretable, computationally fast, and can handle correlated variables. Model utility was examined by (1) accuracy, represented by the area under the receiver operating characteristic curve (AUC) with standard deviation (SD), and (2) explainability, illustrated by the visual examination of the estimated attention maps which showed the important areas for model diagnostics.

RESULTS

During the training and validation stage, we observe that when we provide no guidance from domain knowledge, the IPF diagnosis model reaches acceptable performance (AUC±SD = 0.93±0.07), but lacks explainability; when including only guided high- or medium-resolution attention, the learned attention maps are not satisfactory; when including both high- and medium-resolution attention, under certain hyperparameter settings, the model reaches the highest AUC among all experiments (AUC±SD = 0.99±0.01) and the estimated attention maps concentrate on the regions of interests for this task. Three best-performing hyperparameter selections according to MSGA were applied to the holdout test set and reached comparable model performance to that of the validation set.

CONCLUSIONS

Our results suggest that, for a task with only scan-level labels available, MSGA+RF can utilize the population-level domain knowledge to guide the training of the network, which increases both model accuracy and explainability.

Non-task expert physicians benefit from correct explainable AI advice when reviewing X-rays

Artificial intelligence (AI)-generated clinical advice is becoming more prevalent in healthcare. However, the impact of AI-generated advice on physicians' decision-making is underexplored. In this study, physicians received X-rays with correct diagnostic advice and were asked to make a diagnosis, rate the advice's quality, and judge their own confidence. We manipulated whether the advice came with or without a visual annotation on the X-rays, and whether it was labeled as coming from an AI or a human radiologist. Overall, receiving annotated advice from an AI resulted in the highest diagnostic accuracy. Physicians rated the quality of AI advice higher than human advice. We did not find a strong effect of either manipulation on participants' confidence. The magnitude of the effects varied between task experts and non-task experts, with the latter benefiting considerably from correct explainable AI advice. These findings raise important considerations for the deployment of diagnostic advice in healthcare.