Interpretable multi-model deep learning framework for automated four-class diagnosis of ocular toxoplasmosis using...

Ocular toxoplasmosis (OT) remains a leading cause of posterior uveitis worldwide, where accurate differentiation between Active, Inactive, Mixed, and Healthy retinal states is critical for appropriate management. Existing deep learning studies primarily address binary classification, offering limited diagnostic granularity and interpretability. This study introduces a comprehensive and interpretable multi-model deep learning framework designed for four-class classification of OT using retinal fundus images. The framework integrates convolutional neural networks (CNNs), Vision Transformers (ViT), YOLOv8, YOLOv11, and an ensemble model, optimized through systematic hyperparameter tuning and dataset balancing. Performance was evaluated on multiple dataset variants and externally validated on independent datasets. Interpretability was assessed using SHAP and EigenCAM to ensure transparency and clinical coherence. This framework achieved a 98% classification accuracy and a Matthews Correlation Coefficient (MCC) of 0.885, demonstrating balanced and robust performance across all categories. The YOLOv11 model yielded the highest accuracy, effectively localizing lesion rich regions through anchor-free detection. The Vision Transformer provided superior contextual understanding, while green-channel preprocessing and class balancing significantly enhanced lesion contrast and model generalization. SHAP and EigenCAM analyses confirmed that predictions corresponded to clinically relevant retinal structures, supporting model transparency and interpretability. Inference benchmarking showed the accelerations of Graphics Processing Unit (GPU) and Central Processing Unit (CPU) for real-time clinical deployment, with YOLO v11 achieving a latency of 2.8ms and 65ms respectively. This framework establishes a new benchmark for interpretable AI in ocular toxoplasmosis classification. Its integration of multi-architecture design, spectral preprocessing, and explainable visual reasoning enables accurate, transparent, and generalizable diagnosis. These findings highlight this study's potential for real-world integration into teleophthalmology, screening programs, and clinical decision-support systems, paving the way for trustworthy and scalable retinal AI solutions.