Multimodal Ocular Disease Recognition

Multimodal Ocular Disease Recognition Bc. Daniel Havel Department of Software and Computer Science Education Abstract Fundus photography enables non-invasive detection of vision-threatening conditions, yet manual diagnosis remains resource-intensive due to specialist shortages. This thesis sys- tematically investigates multimodal learning for automated multi-label disease classifica- tion on the ODIR dataset, which provides bilateral fundus images, patient metadata, and diagnostic keywords. We conduct 20 controlled experiments comparing single-eye versus bilateral inputs, demographic and keyword integration, fusion strategies, preprocessing, and sampling policies. Using frozen DINOv2 features as baseline, even a minimal linear classifier achieves a Final Score of 0.725. Our best configuration-late fusion with meta- data, keyword prediction, and partial unfreezing-yields 0.742 Final Score and improves Cohen's Kappa by 5.2% over baseline. We find that aggressive oversampling degrades per- formance, while LAB-space CLAHE and higher resolution provide incremental benefits. Overall, this study clarifies where multimodal integration delivers meaningful improve- ments in clinical classification metrics. Keywords: Multimodal data, Deep Learning, Multi-label classification, Disease Recog- nition, ODIR dataset