Inference Time Adaptation for Improved Retinal Disease Diagnosis Using Optical Coherence Tomography Images

Spectral-Domain Optical Coherence Tomography (SDOCT) is the most common and cost-effective variant of OCT for diagnosing retinal diseases. Deep learning models that are utilized for automated diagnosis of these images often exhibit poor performance - across data-sets with varying acquisition settings, limiting their clinical utility. This work, for the first time proposes “Inference Time Adaptation (ITA)” with lightweight deep learning models for SDOCT analysis, enhancing model deployability and generalizability. Furthermore, this work extends ITA, introducing “Inference Time Adaptation with Domain Transforms (ITADT)” to improve the generalizability of deep learning models for retinal disease diagnosis. Domain specific augmentations are applied to test samples to enable regularized model adaptation, enhancing classification performance. The proposed approach is model agnostic and can be applied to any trained deep learning model. The proposed methods achieve significant performance improvement of > 10% on average in accuracy, compared with the un-adapted/original deep learning model on benchmark data-sets for classification of retinal diseases.