A Deep Learning Model for Detecting the Eyes Receiving Glaucoma Medications Using Anterior Segment Images.

Abstract

Purpose: We aimed to investigate whether a deep learning model can detect eyes receiving glaucoma medications from anterior segment images and to visualize the anatomical areas prioritized during classification.

Methods: The training dataset was comprised of 20,000 augmented images of eyes receiving or not receiving glaucoma medications. The test dataset was comprised of 100 images each of eyes receiving and not receiving glaucoma medications. Diagnostic performance of the model was evaluated using the area under the receiver operating characteristic curve (AROC) and compared with human recognition. Subgroup analyses were performed based on conjunctival hyperemia, prostaglandin analog use, and illumination conditions. Gradient-Weighted Class Activation Mapping (Grad-CAM) was applied to explore anatomical areas prioritized by the model.

Results: The deep learning model detected the eyes receiving glaucoma medications with significantly higher accuracy than human recognition (AROC, 0.90 vs. 0.75; P < 0.01). No significant AROC differences were observed in the presence or absence of conjunctival hyperemia, prostaglandin analog use, or under varying illumination. Grad-CAM analysis revealed the periocular area was significantly more frequently highlighted in eyes receiving glaucoma medication than in those not receiving medication (P < 0.01).

Conclusions: The deep learning model objectively detected glaucoma medication use based on anterior segment images. Saliency mapping suggests that the model can identify subtle periocular changes induced by treatment.

Translational relevance: The deep learning model will contribute to assessing the severity of side-effects of glaucoma medications and facilitate the development of eye drops with improved tolerability.