Deep Learning Algorithm Prognosticating Retinal Tears and Detachments From Optical Coherence Tomography

Our image classifier prognosticates future retinal tear/retinal detachment (RT/RD) likelihood from optical coherence tomography (OCT) while providing pixel-level explanations of clinical importance.

RT/RD status (International Classification of Diseases, Ninth and Tenth Revision codes) and surgical status (Current Procedural Terminology codes) were determined for OCTs collected from the Stanford Research Repository. An image positive for future RT/RD-related surgery was defined as no RT/RD or surgery prior to the acquisition date and the acquisition date 90 days prior to RT/RD diagnosis or surgery. A negative image had no patient overlap with the positive class, had no RT/RD or surgery indication at any time, and was positive for plaquenil use without toxic maculopathy. A convolutional neural network, Inception-v4, was fine-tuned in a class-stratified fivefold fashion on the data set containing 433 negative patients (1027 images) and 343 positive patients (1027 images). Each fold contained a separate patient cohort. Heatmaps indicating a model's region of focus were generated using gradient-weighted class activation mapping to verify that the model's intuition was consistent with clinical knowledge.

Performance metrics were collected by averaging across folds. For the test set, the model achieved an area under the receiver operating characteristic curve of 0.87, an average precision score of 0.85, and an accuracy of 0.78. Anatomy highlighted in heatmaps described macular biomarkers for RT/RD, including epiretinal membrane presence, vitreomacular traction, degree of myopic tilt, and choroidal thickness.

The binary image classifier accurately identified future RT/RD development from OCTs.

Our deep learning algorithm highlights biomarkers for patients at high risk for RT/RDs, providing a window for prophylactic treatment to prevent vision loss. .