Artificial intelligence to enhance the diagnosis of ocular surface squamous neoplasia.

To provide an artificial intelligence (AI) method using in vivo confocal microscopy (IVCM) to differentiate ocular surface squamous neoplasia (OSSN) from other lesions and compare the performance of well-known AI-related solutions. A dataset of 2,774 IVCM images, comprising OSSN and other ocular surface diseases was used to train three deep learning models: ResNet50V2, Yolov8x, and VGG19. These models were trained to identify OSSN-related lesions by recognizing specific visual features, including the "starry-sky" pattern, hyperkeratosis, mitotic figures and irregularly shaped epithelial cells. To mitigate class imbalance, a novel square-based data augmentation strategy was employed. Additionally, we implemented a few-shot learning model to enhance the precision of rare symptoms, such as mitosis. To enhance model interpretation, Shapley values and Uniform Manifold Approximation and Projection (UMAP) analysis were employed to explain decision-making processes. The AI models demonstrated high accuracy in distinguishing healthy tissues from pathological ones, achieving over 90% accuracy across all models. In our binary classification task, all AI models had accuracy above 97% (precision ≥ 98%, recall ≥ 85%, F1 score ≥ 92%). The model achieved lower accuracy in 4 class labeled classification. Aggregation of cell-level results provided the best performance with an F1 score of 100%. The models successfully identified patient-specific features in IVCM images, suggesting that these images can act as "fingerprints". Our AI model utilizing IVCM was able to classify OSSN with high accuracy. Moreover, cell-level classification results could be backpropagated to image-level and patient-level. The patient-specific information within IVCM images offers promise for personalized diagnostics and treatment monitoring in ocular oncology.