An efficient deep learning-based multi-classification of ocular toxoplasmosis and its secondary complications from fundus images

Abstract

"Ocular Toxoplasmosis is considered a leading cause of visual impairment when not diagnosed and treated correctly on time. Its complications pose a challenge for accurate diagnosis and treatment. This study represents an efficient deep-learning approach for the multi-classification of Ocular Toxoplasmosis and its secondary complications using image datasets. In order to obtain a better fit of the model and further improve the data set, we incorporated its secondary complications. By carefully and accurately collecting a diverse set of data, we significantly enriched our custom dataset and enhanced its potential for insightful analysis. Deep learning techniques help us to develop an accurate multi-classification system for ocular toxoplasmosis and its associated complications. In this research, we utilized multiple CNN models and Transformer models, all of which demonstrated excellent accuracy. Additionally, we introduced two hybrid models by combining CNN and Transformer architectures. Finally, we also implemented robust ensemble architecture's two methods by combining the best-performing models: VGG19, ViT, and ResNet50. Furthermore, we integrated the XAI (GRAD-CAM) with the best accuracy 98% and F1 score 97% achiever Hybrid2(VGG19+ViT) model to enhance the understanding of our classification process and provide more transparency in the model’s decision-making."