Deep Learning Classification of Angle Closure based on Anterior Segment OCT

Current methods for identifying angle-closure disease rely on subjective assessments and lack high-throughput capabilities. Prior research has shown that anterior segment OCT provides detailed anatomical data useful for diagnosing eye conditions. However, no prior work had resolved how to apply deep learning models to classify PACD stages automatically. This gap motivated researchers to explore whether machine learning could improve diagnostic accuracy and consistency. While existing tools can detect PACD, they often fail to differentiate between PACS and PACG effectively. The need for objective, scalable methods is clear. Standardized classification systems exist, but their application is limited by human variability. This paper's contribution lies in testing a CNN-based approach for PACD classification. The study's design aims to address limitations in generalizability and diagnostic precision.

The aim of this study was to evaluate a deep learning model for identifying PACD and differentiating its stages using AS-OCT images. Researchers wanted to determine if a CNN could classify PACD with high accuracy and generalizability across different populations. They focused on three classification tasks: distinguishing controls from PACS and PACG, identifying PACD overall, and separating PACS from PACG. The motivation was to develop a tool that could reduce diagnostic variability and increase throughput in clinical settings. The study also aimed to assess how well the model performed when applied to data from a different geographic region. Researchers sought to measure the model's precision and recall across various datasets. The goal was to provide a reliable alternative to subjective diagnostic methods. The study's design allowed for both internal validation and external testing.

The study used a cross-sectional design involving patients from three eye centers in China and Singapore. Participants underwent standardized ophthalmic exams and were classified by physicians into four categories: control, PACS, PAC, or PACG. A CNN model was trained using AS-OCT images to perform three classification tasks. The first classifier separated control eyes from PACS and PACG. The second classifier identified PACD versus controls. The third classifier distinguished PACS from PACG. Training data came from 841 eyes in China, with 300 additional eyes from Singapore used for testing. The model's performance was evaluated using AUC, precision, and recall metrics. Validation sets were tested separately from the training data to ensure accuracy. The study's approach combined clinical data with machine learning techniques.

Classifier 2 achieved the highest generalizability, with an AUC of 0.96 on the validation set and 0.95 on the test set. Classifier 1 had an AUC of 0.96 on the validation set but dropped to 0.84 on the test set. Classifier 3 performed the weakest, with an AUC of 0.83 on validation and 0.64 on testing. These results suggest that the model can reliably identify PACD but struggles with finer stage differentiation. The drop in performance for Classifier 1 indicates challenges in cross-regional generalizability. Classifier 3's poor performance highlights the difficulty in distinguishing PACS from PACG. The model's precision and recall metrics supported these findings. The study's results align with the authors' hypothesis about diagnostic accuracy.

The authors concluded that CNN classifiers can distinguish PACD from controls with good generalizability across different patient cohorts. However, the model's performance is moderate when differentiating PACS from PACG. The results suggest that while deep learning can support diagnostic decisions, it may not replace physician judgment in all cases. The study's findings are based on the authors' stated analysis of the model's performance. No essential role of the model in clinical practice is claimed beyond supporting existing methods. The authors emphasize the importance of further research to improve stage-specific classification. Their conclusions are limited to the data and methods described in the paper. No broader implications beyond PACD classification are proposed.