Ocular toxoplasmosis classification from low-resource dataset leveraging DiNet-fusion

Abstract

Ocular toxoplasmosis (OT) is a vision-threatening retinal illness, and its automated detection is challenging due to the scarcity, imbalance, and heterogeneity of fundus imaging datasets. Beneath these constraints, traditional deep learning approaches are challenged with providing accurate predictions, often resulting in inconsistent, non-clinically relevant multi-class classification between disease classes. This paper introduces DiNet-Fusion, a data-efficient fusion model that integrates EfficientNet- B0 and self-supervised DINO ViT-Small to generate comprehensive local and global representations for multi-class OT classification. Following the cleansing and stratification of a low-resource clinical dataset, features from both backbones were integrated and refined by weighted training and label smoothing to address class imbalance. The findings of the model shows DiNet-Fusion as a robust and resourceefficient technology effective at facilitating OT diagnosis in low-resource dataset clinical environments. Finally, the model demonstrated robust performance, exceeding 92% accuracy on the test set, with threshold optimization enhancing the recall of the clinically significant active class from 78.57% to 92.86%. Additionally, high AUC values (0.97–0.99), consistent learning curves, and assured probability distributions further prove its dependability.