Enhancing Interpretability in Retinal OCT Analysis Using Multi-Channel Feature Encoding.

This study examines the training of lightweight classifiers, specifically ResNet-18 and DenseNet-121, for retinal analysis in optical coherence tomography. Multi-channel inputs were compared against standard grayscale-to-RGB copies, increasing the reliability of the results. While RGB replication is standard for transfer learning, the multi-channel approach enhances feature extraction. Clinical reliability was further validated via Integrated Gradients to audit decision boundaries. This work confirms that specialized input channels help create more interpretable and reliable clinical decision-support systems.