CAJU: Convolutional Attentive Joint Units for Infectious Keratitis In Vivo Confocal Microscopy Image Classification

Infectious keratitis (IK) is a significant cause of preventable vision loss and necessitates expeditious, dependable etiological assessment. Conventional diagnostics (e.g., corneal scraping, culture, and molecular tests) are often invasive, time-consuming, and resource-dependent, which limits their practicality in routine workflows. As a non-invasive alternative, in vivo confocal microscopy (IVCM) enables high-resolution visualization of corneal microstructures and potential pathogens. However, its interpretation requires specialized expertise and remains subject to inter-observer variability. We hereby propose Convolutional Attentive Joint Units (CAJU), an attention-guided dual-stream pipeline for IK classification from IVCM images. CAJU instantiates complementary CNN backbones and fuses their representations through a transformer-inspired self-attention module, allowing backbone selection according to performance and deployment constraints. The public IVCM-Keratitis dataset was subjected to a series of experiments following the implementation of artifact filtration, encompassing a total of 2,155 images classified into four distinct categories: AK, FK, NSK, and Normal. These experiments were executed through the utilization of a cross-validation approach. Among the evaluated instantiations, CAJU (DenseNet-161 + ResNet-101, w/ SA) demonstrated optimal class-balanced performance, attaining 96.22% precision, 96.01% recall, and 96.03% F1-score, while CAJU (VGG16 + MobileNetV3, w/ SA) attained the maximum accuracy of 97.13%. In order to support interpretability, the present study integrates Grad-CAM and SHAP to localize salient regions and provide class-wise attributions, thereby enabling plausibility checks in both correct predictions and clinically relevant error modes (notably NSK ambiguity). In summary, CAJU offers a dependable and comprehensible framework to facilitate IVCM-based IK diagnosis and to support clinical decision-making, particularly in non-specialized settings.