Computer-aided recognition of myopic tilted optic disc using deep learning algorithms in fundus photography

The role of saliency maps in enhancing ophthalmologists' trust in artificial intelligence models

PURPOSE

Saliency maps (SM) allow clinicians to better understand the opaque decision-making process in artificial intelligence (AI) models by visualising the important features responsible for predictions. This ultimately improves interpretability and confidence. In this work, we review the use case for SMs, exploring their impact on clinicians' understanding and trust in AI models. We use the following ophthalmic conditions as examples: (1) glaucoma, (2) myopia, (3) age-related macular degeneration, and (4) diabetic retinopathy.

METHOD

A multi-field search on MEDLINE, Embase, and Web of Science was conducted using specific keywords. Only studies on the use of SMs in glaucoma, myopia, AMD, or DR were considered for inclusion.

RESULTS

Findings reveal that SMs are often used to validate AI models and advocate for their adoption, potentially leading to biased claims. Overlooking the technical limitations of SMs, and the conductance of superficial assessments of their quality and relevance, was discerned. Uncertainties persist regarding the role of saliency maps in building trust in AI. It is crucial to enhance understanding of SMs' technical constraints and improve evaluation of their quality, impact, and suitability for specific tasks. Establishing a standardised framework for selecting and assessing SMs, as well as exploring their relationship with other reliability sources (e.g. safety and generalisability), is essential for enhancing clinicians' trust in AI.

CONCLUSION

We conclude that SMs are not beneficial for interpretability and trust-building purposes in their current forms. Instead, SMs may confer benefits to model debugging, model performance enhancement, and hypothesis testing (e.g. novel biomarkers).

Self-supervised pre-training for joint optic disc and cup segmentation via attention-aware network

Image segmentation is a fundamental task in deep learning, which is able to analyse the essence of the images for further development. However, for the supervised learning segmentation method, collecting pixel-level labels is very time-consuming and labour-intensive. In the medical image processing area for optic disc and cup segmentation, we consider there are two challenging problems that remain unsolved. One is how to design an efficient network to capture the global field of the medical image and execute fast in real applications. The other is how to train the deep segmentation network using a few training data due to some medical privacy issues. In this paper, to conquer such issues, we first design a novel attention-aware segmentation model equipped with the multi-scale attention module in the pyramid structure-like encoder-decoder network, which can efficiently learn the global semantics and the long-range dependencies of the input images. Furthermore, we also inject the prior knowledge that the optic cup lies inside the optic disc by a novel loss function. Then, we propose a self-supervised contrastive learning method for optic disc and cup segmentation. The unsupervised feature representation is learned by matching an encoded query to a dictionary of encoded keys using a contrastive technique. Finetuning the pre-trained model using the proposed loss function can help achieve good performance for the task. To validate the effectiveness of the proposed method, extensive systemic evaluations on different public challenging optic disc and cup benchmarks, including DRISHTI-GS and REFUGE datasets demonstrate the superiority of the proposed method, which can achieve new state-of-the-art performance approaching 0.9801 and 0.9087 F1 score respectively while gaining 0.9657 D C disc and 0.8976 D C cup . The code will be made publicly available.

Deep learning-based optic disc classification is affected by optic-disc tilt

We aimed to determine the effect of optic disc tilt on deep learning-based optic disc classification. A total of 2507 fundus photographs were acquired from 2236 eyes of 1809 subjects (mean age of 46 years; 53% men). Among all photographs, 1010 (40.3%) had tilted optic discs. Image annotation was performed to label pathologic changes of the optic disc (normal, glaucomatous optic disc changes, disc swelling, and disc pallor). Deep learning-based classification modeling was implemented to develop optic-disc appearance classification models with the photographs of all subjects and those with and without tilted optic discs. Regardless of deep learning algorithms, the classification models showed better overall performance when developed based on data from subjects with non-tilted discs (AUC, 0.988 ± 0.002, 0.991 ± 0.003, and 0.986 ± 0.003 for VGG16, VGG19, and DenseNet121, respectively) than when developed based on data with tilted discs (AUC, 0.924 ± 0.046, 0.928 ± 0.017, and 0.935 ± 0.008). In classification of each pathologic change, non-tilted disc models had better sensitivity and specificity than the tilted disc models. The optic disc appearance classification models developed based all-subject data demonstrated lower accuracy in patients with the appearance of tilted discs than in those with non-tilted discs. Our findings suggested the need to identify and adjust for the effect of optic disc tilt on the optic disc classification algorithm in future development.

Quantization of Optic Disc Characteristics in Young Adults Based on Artificial Intelligence

PURPOSE

This study aimed to automatically and quantitatively analyse the characteristics of the optic disc by applying artificial intelligence (AI) to fundus images.

METHODS

A total of 1084 undergraduates were recruited in this cross-sectional study. The optic disc area, cup-to-disc ratio (C/D), optic disc tilt, and the area, width, and height of peripapillary atrophy (PPA) were automatically and quantitatively detected using AI. Based on axial length (AL), participants were divided into five groups: Group 1 (AL ≤ 23 mm); Group 2 (23 mm < AL≤ 24 mm); Group 3 (24 mm < AL≤ 25 mm); Group 4 (25 mm < AL< 26 mm) and Group 5 (AL ≥ 26 mm). Relationships between ocular parameters and optic disc characteristics were analysed.

RESULT

A total of 999 undergraduates were included in the analysis. The prevalence of optic disc tilting and PPA were 47.1% and 92.5%, respectively, and increased with the severity of myopia. The mean optic disc area, PPA area, C/D, and optic disc tilt ratio were 1.97 ± 0.46 mm2, 0.84 ± 0.59 mm2, 0.18 ± 0.07, and 0.81 ± 0.08, respectively. In Group 5, the average optic disc area (1.84 ± 0.41 mm2) and optic disc tilt ratio (0.79 ± 0.08) were significantly smaller and the PPA area (1.12 ± 0.61 mm2) was significantly larger than those in the other groups. AL was negatively correlated with optic disc area and optic disc tilt ratio (r=-0.271, -0.219; both p < 0.001) and positively correlated with PPA area, width, and height (r = 0.421, 0.426, 0.345; all p < 0.01). A greater AL (β = 0.284, p < 0.01) and a smaller optic disc tilt ratio (β=-0.516, p < 0.01) were related to a larger PPA area.

CONCLUSION

The characteristics of the optic disc can be feasibly and efficiently extracted using AI. The quantization of the optic disc might provide new indicators for clinicians to evaluate the degree of myopia.

Application of Machine Learning Algorithm in Predicting Axillary Lymph Node Metastasis from Breast Cancer on Preoperative Chest CT

Axillary lymph node (ALN) status is one of the most critical prognostic factors in patients with breast cancer. However, ALN evaluation with contrast-enhanced CT (CECT) has been challenging. Machine learning (ML) is known to show excellent performance in image recognition tasks. The purpose of our study was to evaluate the performance of the ML algorithm for predicting ALN metastasis by combining preoperative CECT features of both ALN and primary tumor. This was a retrospective single-institutional study of a total of 266 patients with breast cancer who underwent preoperative chest CECT. Random forest (RF), extreme gradient boosting (XGBoost), and neural network (NN) algorithms were used. Statistical analysis and recursive feature elimination (RFE) were adopted as feature selection for ML. The best ML-based ALN prediction model for breast cancer was NN with RFE, which achieved an AUROC of 0.76 ± 0.11 and an accuracy of 0.74 ± 0.12. By comparing NN with RFE model performance with and without ALN features from CECT, NN with RFE model with ALN features showed better performance at all performance evaluations, which indicated the effect of ALN features. Through our study, we were able to demonstrate that the ML algorithm could effectively predict the final diagnosis of ALN metastases from CECT images of the primary tumor and ALN. This suggests that ML has the potential to differentiate between benign and malignant ALNs.

Semi-supervised automatic seizure detection using personalized anomaly detecting variational autoencoder with behind-the-ear EEG

BACKGROUND AND OBJECTIVE

Epilepsy is one of the most common neurologic diseases worldwide, and 30% of the patients live with uncontrolled seizures. For the safety of patients with epilepsy, an automatic seizure detection algorithm for continuous seizure monitoring in daily life is important to reduce risks related to seizures, including sudden unexpected death. Previous researchers applied machine learning to detect seizures with EEG, but the epileptic EEG waveform contains subtle changes that are difficult to identify. Furthermore, the imbalance problem due to the small proportion of ictal events caused poor prediction performance in supervised learning approaches. This study aimed to present a personalized deep learning-based anomaly detection algorithm for seizure monitoring with behind-the-ear electroencephalogram (EEG) signals.

METHODS

We collected behind-the-ear EEG signals from 16 patients with epilepsy in the hospital and used them to develop and evaluate seizure detection algorithms. We modified the variational autoencoder network to learn the latent representation of normal EEG signals and performed seizure detection by measuring the anomalies in EEG signals using the trained network. To personalize the algorithm, we also proposed a method to calibrate the anomaly score for each patient by comparing the representations in the latent space.

RESULTS

Our proposed algorithm showed a sensitivity of 90.4% with a false alarm rate of 0.83 per hour without personal calibration. On the other hand, the one-class support vector machine only showed a sensitivity of 84.6% with a false alarm rate of 2.17 per hour. Furthermore, our proposed model with personal calibration achieved 94.2% sensitivity with a false alarm rate of 0.29 while detecting 49 of 52 ictal events.

CONCLUSIONS

We proposed a novel seizure detection algorithm with behind-the-ear EEG signals via semi-supervised learning of an anomaly detecting variational autoencoder and personalization method of anomaly scoring by comparing latent representations. Our approach achieved improved seizure detection with high sensitivity and a lower false alarm rate.

Deep learning models for screening of high myopia using optical coherence tomography

This study aimed to validate and evaluate deep learning (DL) models for screening of high myopia using spectral-domain optical coherence tomography (OCT). This retrospective cross-sectional study included 690 eyes in 492 patients with OCT images and axial length measurement. Eyes were divided into three groups based on axial length: a “normal group,” a “high myopia group,” and an “other retinal disease” group. The researchers trained and validated three DL models to classify the three groups based on horizontal and vertical OCT images of the 600 eyes. For evaluation, OCT images of 90 eyes were used. Diagnostic agreements of human doctors and DL models were analyzed. The area under the receiver operating characteristic curve of the three DL models was evaluated. Absolute agreement of retina specialists was 99.11% (range: 97.78–100%). Absolute agreement of the DL models with multiple-column model was 100.0% (ResNet 50), 90.0% (Inception V3), and 72.22% (VGG 16). Areas under the receiver operating characteristic curves of the DL models with multiple-column model were 0.99 (ResNet 50), 0.97 (Inception V3), and 0.86 (VGG 16). The DL model based on ResNet 50 showed comparable diagnostic performance with retinal specialists. The DL model using OCT images demonstrated reliable diagnostic performance to identify high myopia.

ODTiD: Optic Nerve Head SD-OCT Image Dataset

Introduction

Optic disc tilt (ODT) or tilted optic disc is a common finding in the general population. It is due to anomalous development caused by the malclosure of the embryonic optic fissure. ODT is commonly associated with high myopia as well as other conditions. In recent days, the common method to image the optic disc (OD) is by optical coherence tomography (OCT). To the best of our knowledge, there are no datasets of ODT available in the public domain. This dataset aims to make open access raw ODT OCT images to test out new image processing segmentation algorithms.

Methods

This dataset of ODT images contains both horizontal and vertical cross-sectional images obtained using spectral-domain optical coherence tomography (SD-OCT, Cirrus 5000, Carl Zeiss Meditec Inc., Dublin, CA). The optic disc cube 200×200 program was used and all the images are aligned with the center of the optic nerve head. This dataset includes images from both clinically normal (20 eyes) and myopic subjects (101 eyes).

Results

The dataset consists of clear (121) and manually marked (121) images resulting in a total of 242 images. The age distribution for all subjects combined is 27.24 ± 9.28 (range, 11.0-69.0) years. For normal subjects mean ± SD age distribution is 32.40 ± 17.23 years. Similarly, the myopia age distribution is 26.22 ± 6.37 years. Ground truth images, ie, manually segmented by a clinical expert are provided along with other meta-data includes age, gender, laterality, refractive error classification, spherical equivalent (SE), best-corrected visual acuity (BCVA), intraocular pressure (IOP), and axial length (AXL).

Conclusion

This open, public database is online at the ICPSR website of the University of Michigan. The dataset can be used to test and validate newly developed automated segmentation algorithms.

Diagnosis of central serous chorioretinopathy by deep learning analysis of en face images of choroidal vasculature: A pilot study

Purpose

To diagnose central serous chorioretinopathy (CSC) by deep learning (DL) analyses of en face images of the choroidal vasculature obtained by optical coherence tomography (OCT) and to analyze the regions of interest for the DL from heatmaps.

Methods

One-hundred eyes were studied; 53 eyes with CSC and 47 normal eyes. Volume scans of 12×12 mm square were obtained at the same time as the OCT angiographic (OCTA) scans (Plex Elite 9000 Swept-Source OCT^®, Zeiss). High-quality en face images of the choroidal vasculature of the segmentation slab of one-half of the subfoveal choroidal thickness were created for the analyses. The 100 en face images were divided into 80 for training and 20 for validation. Thus, we divided it into five groups of 20 eyes each, trained the remaining 80 eyes in each group, and then calculated the correct answer rate for each group by validation with 20 eyes. The Neural Network Console (NNC) developed by Sony and the Keras-Tensorflow backend developed by Google were used as the software for the classification with 16 layers of convolutional neural networks. The active region of the heatmap based on the feature quantity extracted by DL was also evaluated as the percentages with gradient-weighted class activation mapping implemented in Keras.

Results

The mean accuracy rate of the validation was 95% for NNC and 88% for Keras. This difference was not significant (P >0.1). The mean active region in the heatmap image was 12.5% in CSC eyes which was significantly lower than the 79.8% in normal eyes (P<0.01).

Conclusions

CSC can be automatically diagnosed by DL with high accuracy from en face images of the choroidal vasculature with different programs, convolutional layer structures, and small data sets. Heatmap analyses showed that the DL focused on the area occupied by the choroidal vessels and their uniformity. We conclude that DL can help in the diagnosis of CSC.

A method for the automatic detection of myopia in Optos fundus images based on deep learning

Myopia detection is significant for preventing irreversible visual impairment and diagnosing myopic retinopathy. To improve the detection efficiency and accuracy, a Myopia Detection Network (MDNet) that combines the advantages of dense connection and Residual Squeeze-and-Excitation attention is proposed in this paper to automatically detect myopia in Optos fundus images. First, an automatic optic disc recognition method is applied to extract the Regions of Interest and remove the noise disturbances; then, data augmentation techniques are implemented to enlarge the data set and prevent overfitting; moreover, an MDNet composed of Attention Dense blocks is constructed to detect myopia in Optos fundus images. The results show that the Mean Absolute Error of the Spherical Equivalent detected by this network can reach 1.1150 D (diopter), which verifies the feasibility and applicability of this method for the automatic detection of myopia in Optos fundus images.

Generalisability through local validation: overcoming barriers due to data disparity in healthcare

Cho et al. report deep learning model accuracy for tilted myopic disc detection in a South Korean population. Here we explore the importance of generalisability of machine learning (ML) in healthcare, and we emphasise that recurrent underrepresentation of data-poor regions may inadvertently perpetuate global health inequity.

Creating meaningful ML systems is contingent on understanding how, when, and why different ML models work in different settings. While we echo the need for the diversification of ML datasets, such a worthy effort would take time and does not obviate uses of presently available datasets if conclusions are validated and re-calibrated for different groups prior to implementation.

The importance of external ML model validation on diverse populations should be highlighted where possible – especially for models built with single-centre data.