Deep Learning Classification of Uveal Melanoma and Metastatic Carcinoma on Cytology Specimens: A Single-Center Experience

Purpose: Differentiating between a choroidal melanoma and metastatic carcinoma to the choroid on cytological specimens presents significant diagnostic challenges due to subtle morphological differences in the sparse cells available in the sample and the increasing scarcity of specialist ocular pathologists. This study developed and evaluated deep learning models for automated classification of May-Grünwald-Giemsa–stained cytospin whole slide images (WSIs) to support diagnostic decision-making. Methods: Cytospin WSIs from 87 patients (50 with choroidal melanoma and 37 with metastatic carcinoma to the choroid) who underwent trans-scleral fine-needle aspiration biopsy at the Liverpool Ocular Oncology Center were digitized and tiled into 256 × 256 pixel patches. Variants of U-Net architectures were trained on 692 manually annotated tiles for tumor cell segmentation. The diagnostic decision model was developed and evaluated using different classification frameworks for best diagnostic performance: five convolutional neural network, three vision transformer (ViT) models, and seven multiple instance learning (MIL) architectures, with and without three preprocessing approaches (original, segmented, and mask-out tiles). Performance was assessed using three-fold patient-level cross-validation, with 20% of the cohort (n = 17 of the 87 patients) held out as a final independent test set. Results: The U-Net achieved strong segmentation performance with a mean absolute error of 1.91 ± 1.31 µm, precision of 0.80 ± 0.17, and recall of 0.76 ± 0.20. Double-tier feature distillation–MIL with mask-out tiles and a fine-tuned Swin-Tiny/4 back-bone showed the highest patient-level area under the curve (AUC) of all models (AUC, 0.96; 95% confidence interval, 0.83–1.00), while DenseNet121 had the highest patient-level accuracy (0.84). Convolutional neural network models performed more consistently than other frameworks over all preprocessing methods, while MIL architectures with domain-specific features achieved the strongest overall AUC on mask-out tiles. ViTs varied somewhat over the different preprocessing techniques with mask-out tiles achieving the highest ViT performance (Swin-Tiny/4; AUC, 0.87). Conclusions: This single-center experience demonstrates the application of deep learning to cytospin WSIs in ocular oncology, establishing a complete pipeline from tumor cell segmentation to slide-level classification. The framework provides a foundation for artificial intelligence–assisted cytopathology that could accelerate diagnosis, particularly in centers lacking specialist ocular pathology expertise, and may be adaptable to other rare cytological specimens.