A feature agnostic approach for glaucoma detection in OCT volumes (1807.04855v4)

Abstract: 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.

• The paper introduces a feature-agnostic 3D CNN that analyzes raw OCT volumes, eliminating the need for segmentation-based feature engineering.
• It achieves an AUC of 0.94, outperforming classical methods such as Logistic Regression with segmentation features (AUC of 0.89).
• The model incorporates Global Average Pooling to generate Class Activation Maps, enhancing interpretability by highlighting key glaucoma markers.

A Feature Agnostic Approach for Glaucoma Detection in OCT Volumes

The paper presents a novel approach for glaucoma detection employing deep learning, specifically using a 3D Convolutional Neural Network (CNN) to classify optical coherence tomography (OCT) volumes. Historically, glaucoma detection from OCT data relied heavily on segmentation-based features, which demand the delineation of specific anatomical markers such as retinal nerve fiber layer (RNFL) thickness and other structural characteristics of the optic nerve head (ONH). However, this paper proposes a paradigm shift by eliminating the need for feature engineering, thus enabling direct analysis of raw OCT data, avoiding laborious segmentation processes.

Methodology and Results

The authors utilized a database containing 1110 OCT scans, out of which 847 were eyes diagnosed with primary open-angle glaucoma (POAG) and 263 healthy controls. These scans were acquired using a Cirrus SD-OCT Scanner. The OCT volumes were downsampled from their original resolution to 64x64x128 voxels due to GPU memory constraints, allowing the 3D CNN to be trained efficiently.

The proposed CNN architecture consists of five 3D convolutional layers with ReLU activations and batch normalization, employing Global Average Pooling (GAP), which facilitates the generation of Class Activation Maps (CAMs). This is a significant advantage, as it visualizes regions in the OCT volume that are critical for glaucoma identification, such as neuroretinal rim and optic disc cupping.

The paper compared the performance of various machine learning models using segmentation-based features against the feature-agnostic 3D CNN. Classical methods, including Logistic Regression, Support Vector Machine (SVM), and Random Forest, were assessed using a predefined set of 22 anatomical measurements. The highest classification accuracy among these was achieved by Logistic Regression with an Area Under the Receiver Operator Characteristic Curve (AUC) of 0.89. Conversely, the deep learning model achieved a superior performance with an AUC of 0.94, underscoring its effectiveness over feature-dependent methodologies without the requirement of explicit segmentation.

Implications and Future Directions

This paradigm shift to feature-agnostic analysis in glaucoma detection reflects a broader trend in AI research, where raw data can now be directly leveraged by deep learning models to improve both efficiency and accuracy. The ability to generate CAMs further emphasizes the interpretability of deep learning models, offering insights into potential new markers of glaucoma within OCT data, such as changes discernible in the lamina cribrosa, a region gaining attention for its relevance in glaucoma progression.

The paper offers groundwork for the expansion of similar approaches to other OCT scans beyond the optic nerve head, such as the macula, and integration with clinical metadata like IOP and visual field measurements—a direction ripe with potential for enhancing glaucoma diagnostic precision. Moreover, further refinement in network architectures, such as deeper networks with larger datasets and advanced regularization techniques, is anticipated to boost classification outcomes even further.

In conclusion, the presented feature-agnostic approach advances the utility of OCT data in the clinical assessment of glaucoma, reducing dependency on feature-engineering while maintaining high diagnostic accuracy. This approach highlights the transformative potential of deep learning in medical imaging, paving the way for more robust, scalable solutions in disease detection and management.