CSwinFK: a bidirectional cross-attention CNN-Swin Transformer fusion model for fungal keratitis classification in in vivo confocal microscopy images

Fungal keratitis is a corneal infection caused by filamentous fungi or yeasts. It can progress rapidly and, if not diagnosed and treated promptly, may lead to irreversible visual impairment or blindness. In this study, we developed CSwinFK, a classification model for distinguishing fungal keratitis from negative cases using in vivo confocal microscopy images. This model is based on a novel bidirectional cross-attention fusion of the CNN architecture and Swin Transformer. Through deep interactions between the two branches, the model aims to achieve high sensitivity to local structural details while effectively capturing long-range global dependencies in the image. Experiments were conducted on a curated set of 3,612 clinical IVCM images from the fungal keratitis and negative groups using a subject-level data split. The model achieved a test accuracy of 97.23%, with a sensitivity of 97.05% and a specificity of 97.42%. Comparative experiments with representative architectures and existing deep learning models showed that CSwinFK achieved favorable classification performance. Grad-CAM analysis further showed that the model focused on regions exhibiting patterns consistent with fungal hyphae in IVCM images.