Source-linked AI summary
A feature agnostic approach for glaucoma detection in OCT volumes
Stefan Maetschke, Bhavna Antony, Hiroshi Ishikawa, Gadi Wollstein, Joel S. Schuman, Rahil Garnavi
TL;DR
Glaucoma detection traditionally relies on manually designed features extracted from segmented OCT images, whose accuracy can be constrained when segmentation is difficult. This work classifies glaucoma directly from raw ONH-centered OCT volumes using a 3D CNN and compares it with feature-based methods, finding comparable or better detection while identifying informative OCT regions.
Problem
Traditional glaucoma classifiers rely on manually designed features from segmented OCT layers, but accurate segmentation can be difficult in advanced glaucoma, low-quality scans, and co-existing retinal disease.
Method
A 3D CNN classifies glaucomatous eyes directly from unprocessed, ONH-centered OCT volumes and is compared with classical methods trained on segmentation-based features.
Results
The raw-OCT approach achieved detection accuracy comparable or better than traditional feature-based approaches and identified OCT regions important for glaucoma classification using CAMs.
Takeaways & Limitations
CAMs may support discovery of novel or more robust glaucoma disease markers while extending detection to scans that are difficult to segment accurately.
Takeaways & Limitations
The evaluation was limited to ONH scans; adding macula scans, intraocular pressure, and visual-test measurements may further increase accuracy.
Abstract
from arXiv · showhide
Optical coherence tomography (OCT) based measurements of retinal layer thickness, such as the retinal nerve fibre layer (RNFL) and the ganglion cell with inner plexiform layer (GCIPL) are commonly used for the diagnosis and monitoring of glaucoma. Previously, machine learning techniques have utilized segmentation-based imaging features such as the peripapillary RNFL thickness and the cup-to-disc ratio. Here, we propose a deep learning technique that classifies eyes as healthy or glaucomatous directly from raw, unsegmented OCT volumes of the optic nerve head (ONH) using a 3D Convolutional Neural Network (CNN). We compared the accuracy of this technique with various feature-based machine learning algorithms and demonstrated the superiority of the proposed deep learning based method. Logistic regression was found to be the best performing classical machine learning technique with an AUC of 0.89. In direct comparison, the deep learning approach achieved a substantially higher AUC of 0.94 with the additional advantage of providing insight into which regions of an OCT volume are important for glaucoma detection. Computing Class Activation Maps (CAM), we found that the CNN identified neuroretinal rim and optic disc cupping as well as the lamina cribrosa (LC) and its surrounding areas as the regions significantly associated with the glaucoma classification. These regions anatomically correspond to the well established and commonly used clinical markers for glaucoma diagnosis such as increased cup volume, cup diameter, and neuroretinal rim thinning at the superior and inferior segments.
1. Introduction
The paper motivates glaucoma detection from OCT and proposes classifying raw, unsegmented ONH volumes with a 3D CNN instead of relying on manually extracted segmentation-based features. It compares this feature-agnostic approach with classical machine learning trained on traditional OCT measurements.
- Glaucoma affects the optic nerve and is characterized by progressive thinning of the neuroretinal rim.
- OCT provides three-dimensional retinal imaging and clinically useful RNFL and GCIPL biomarkers that thin as glaucoma progresses.
- Classical machine learning commonly uses segmented OCT features such as peripapillary RNFL thickness and macular GCIPL thickness.
- Prior feature-based studies reported peak AUCs from 0.946 to 0.991 using extracted OCT, clinical, or visual-field features.
- Segmentation-based approaches can be difficult in advanced glaucoma, low-quality scans, and co-existing retinal pathologies, while human-selected markers may limit accuracy.
- The proposed CNN detects glaucomatous eyes directly from unprocessed ONH-centered OCT volumes without segmentation, visual fields, or healthy-population thickness profiles.
2. Material and methods
The study evaluates feature-based classifiers and a 3D CNN on ONH-centered OCT scans, using patient-separated data splits and AUC-based assessment. The CNN receives minimally preprocessed volumes and is designed to produce three-dimensional Class Activation Maps.
- 2.1. Performance metric: Performance was measured with AUC, using false-positive and true-positive rates from ranked classifier confidence scores.
- 2.2. Data: The dataset contains 1,110 ONH-centered OCT scans from 624 patients, including 263 healthy and 847 POAG scans.Scans with signal strength less than 7 were discarded.
- 2.2. Data: The data were divided into 888 training, 112 validation, and 110 test samples without splitting eyes from the same patient across folds.The study also performed 5-fold cross-validation.
- 2.3. Feature-based approach: Feature-based models used 22 normalized scanner measurements, including RNFL thickness, rim and disc areas, cup-to-disc ratios, and cup volume.
- 2.3. Feature-based approach: The study trained Naive Bayes, logistic regression, SVM, Random Forest, Gradient Boosting, and Extra Trees classifiers on the extracted features.
- 2.4. Feature agnostic approach: The feature-agnostic model used raw OCT volumes down-sampled from 200x200x1024 to 64x64x128, with no other preprocessing or data extraction.Down-sampling was performed because of 12GB GPU-memory constraints.
- 2.4. Feature agnostic approach: The CNN comprises five 3D convolutional layers followed by global average pooling and dense softmax output for class prediction and CAM computation.3D convolutions preserve depth information and enable three-dimensional CAMs.
3. Results
The results compare feature-based classifiers with a feature-agnostic CNN and examine which OCT regions and classical features contribute to glaucoma classification.
- Disease detection: 0.89 AUC was achieved by logistic regression, the best classical feature-based method on the test data.Tree-based classifiers tended to overfit, with small validation–test differences observed for lower-capacity classifiers.
- Feature importance: The Extra Trees analysis found the largest feature importance for 6 and 11 o’clock clock-hours, inferior and superior quadrants, and vertical cup-to-disc ratio.Feature importance was estimated over 100 folds with standard deviations shown in Figure 2.
- Disease detection: 0.94 peak test AUC was achieved by the feature-agnostic CNN, substantially exceeding the best classical method’s 0.89 AUC.The CNN was evaluated using 5-fold cross-validation on the OCT dataset.
- CNN attention: CAMs for POAG eyes highlighted optic disc cupping, neuroretinal rims, the lamina cribrosa, and surrounding regions.These regions correspond to established clinical markers including cup diameter or volume and rim area or volume.
4. Discussion
The discussion contrasts raw-volume 3D CNN classification with feature-based approaches, interprets CAM-highlighted anatomy, and notes architectural, dataset, and cohort limitations.
- Comparison with prior work: Classical approaches reported peak AUCs of 0.95–0.99, whereas this study measured a test AUC of 0.89 for classical algorithms on its dataset.Earlier studies often used smaller datasets, non-hold-out evaluation, feature selection, and additional clinical measurements.
- CAM interpretation: The 3D CNN’s CAMs highlighted neuroretinal rim thinning, optic disc cupping, and the lamina cribrosa with surrounding areas.These regions correspond to established glaucoma markers and may support investigation of additional biomarkers.
- CAM interpretation: CAM resolution depends on the number of preceding pooling or strided-convolution operations, so aggressive downsampling can produce very small activation maps.An input volume of 128x128x128 becomes 4x4x4 after five stride-2 reductions.
- Architectural limitations: GPU memory limits forced a comparatively shallow five-layer network, although deeper or differently designed networks might achieve higher classification accuracy.The available memory was 12GB.
- Cohort limitations: Age differed significantly between groups and was not directly included as a feature, but the CNN could potentially infer age-related information from OCT layer thinning.Differences in IOP, MD, and GHT were expected between healthy and glaucomatous groups.
5. Conclusions
The study shows that glaucoma detection from raw OCT volumes can match or exceed segmentation-based approaches. CAMs also identify relevant OCT regions, while the evaluation remains limited to ONH scans.
- Conclusions: Raw OCT-volume detection achieved accuracy comparable to or better than traditional approaches using manually designed features from segmented OCTs.The feature-agnostic design may be useful when segmentation is difficult in older patients or extreme glaucoma cases.
- Conclusions: CAMs identified OCT regions important for glaucoma classification and may help discover novel or more robust disease markers.This provides an interpretability function analogous to the human interpretability of manually designed features.
- Scope: The results were based on the largest OCT glaucoma dataset reported by the authors but were limited to ONH scans only.Adding macular scans, IOP, and visual-test measurements might further increase accuracy.