Explainable Neutrosophic Knowledge Distillation Model for Ocular Disease Classification Using Ultra-Wide Field Fundus Images

Abstract

Ultra-wide field (UWF) fundus image classification is an important part of the entire process of medical screening and decision support. However, the discrimination of various retinal disease classes is difficult due to the similarity between classes, class imbalance, and the indeterminacy of visual patterns. In our research, an explainable neutrosophic knowledge distillation (NKD) model for UWF fundus image classification is proposed. In the proposed model, the teacher model is a ResNet50 architecture that provides the student model with supervisory information that is aware of the indeterminacy of predictions. The proposed model combines the CLAHE-based preprocessing method with the neutrosophic distillation method to enable the student model to learn from the hard labels as well as the teacher model. The experimental results were evaluated using the 5-fold cross-validation method with an additional hold-out evaluation. The experimental results show that the proposed NKD model has a mean accuracy of 84.00%, specificity of 97.33%, precision of 84.99%, recall of 84.00%, and F1-score of 84.02%. The proposed model also has an accuracy of 87.86% with specificity of 97.48% and AUC of 97.48% in the ablation-based full model evaluation. It outperformed classical machine learning baselines based on Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG), and LBP + HOG features with Support Vector Machines (SVM) classifiers, as well as the baseline student, fuzzy student, and teacher Convolutional Neural Network (CNN) models. For improved interpretability, the Grad-CAM++ technique was used to analyze the proposed NKD model. This analysis showed that the network attended to relevant retinal regions during classification. These results suggest that the proposed model can be an effective tool for UWF fundus image classification.