A Deep Learning Framework with Qualitative Post Hoc Explanation for Dry Eye Detection and Tear Film Lipid Layer Pattern Classification Using Anterior-Segment and Tearscope Imaging

Background/Objectives: Dry eye disease (DED) is assessed by combining patient-reported symptoms with ocular surface findings. We evaluated a deep learning strategy with qualitative post hoc explanation in two independent ophthalmic imaging tasks: DED-versus-healthy classification from anterior segment slit lamp images; and four-class recognition of tear film lipid layer (TFLL) interference patterns in Tearscope images using the Guillon categories. Methods: ResNet backbones were modified by adding parallel expansion channels while keeping the ImageNet-initialized channels frozen. The added channels were He-normal initialized and trained at 1 × 10−4, with the MLP classification head trained at 1 × 10−3. Latin hypercube sampling was used for hyperparameter selection within the development data. We used Grad-CAM and Grad-CAM++ only for post hoc interpretation. We evaluated ShECS slit lamp images (n = 434) and VOPTICAL-Is Tearscope images (n = 406). Results: Under the original evaluations, ResNet-101×3 achieved the highest scores among the evaluated configurations on both tasks. On ShECS, accuracy was 74% (95% CI: 0.63–0.82), with precision 0.80, recall 0.90, and F1-score 0.85 for the DED (positive) class, and AUC 0.80. On VOPTICAL-Is, the original image-level evaluation yielded four-class TFLL accuracy of 90% (95% CI: 0.82–0.95), with macro precision 0.89, macro recall 0.90, and macro F14 0.89, and a macro one-versus-rest AUC of 0.98. Conclusions: These preliminary proof-of-concept results support further evaluation of the combined channel-expansion and differential-training strategy, together with qualitative post hoc explanation methods for ophthalmic image classification. Because VOPTICAL-Is contains only healthy-subject images and has no documented subject identifiers in the accessed release (group-code overlap between the development and held-out subsets was observed, and subject overlap cannot be excluded), the second task represents image-level TFLL pattern classification rather than clinical DED-subtype diagnosis. A separate code-grouped sensitivity analysis with a frozen ResNet-50 feature extractor yielded 0.59 held-out accuracy (macro one-versus-rest AUC 0.85). Because this grouped run did not re-run the 0.90 ResNet-101×3 configuration, the numerical difference cannot be attributed solely to partitioning, or interpreted as a direct measurement of subject-level leakage; rather, it demonstrates sensitivity to group-aware partitioning. The original image-level estimates should therefore be treated as exploratory and not as evidence of subject-level generalization. A controlled 2 × 2 factorial ablation (channel width × trainability policy) on the same code-grouped partition, executed as a TensorFlow/Keras ResNet-50 reconstruction with three seeds per cell, indicated descriptively that unfreezing the final convolutional stage was associated with higher accuracy and macro F1 (e.g., standard width: accuracy 0.556 vs. 0.469), whereas the effect of widening was small and inconsistent across metrics and policies. These descriptive summaries involve no significance testing and do not measure subject-level leakage. External clinical validation of identified cohorts is needed before routine use.