Enhancing cataract diagnosis: towards fine-grained cataract classification using deep learning and progressive image processing strategies

Abstract

The sense of seeing the world is one the most precious gifts of human life. However, cataract remains a major public health threat because of its high prevalence and frequent causing severe vision impairment or even blindness. Early diagnosis continues to be critical but reliance on ophthalmologists restricts availability, particularly in low-resourced areas. To solve this problem, in this paper, we present an AIassisted fine-grained multiclass cataract classification framework, based on effective deep learning and progressive image preprocessing techniques. Unlike other prior works that only used public datasets, we assembled a real-world anterior eye image dataset acquired from hospitals in the city of Dhaka (Bangladesh), and annotated by experienced ophthalmologists. Our model is based on a variety of deep learning architectures that include VGG19, ResNet101, InceptionV3, Xception, EfficientNetB2, DenseNet121, MobileNetV2, FastViT, MobileViT, and ViT_B16, which were finetuned, using transfer learning, for the task of multiclass classification with four different cataract types as well as normal cases. A full preprocessing pipeline that includes resizing, normalization, augmentation and contrast enhancement was applied to enhance the generalization capability among the models. For further boost in performance, we utilized two ensemble methods, bagging and stacking, where bagging combines the predictions of those top-performing models via averaging the probabilities for final decision, whereas a stacking meta-learner (logistic regression) combines the base models output for improved estimation. Both systematic ensembler approaches yielded striking improvements in accuracy and performance across varied test conditions. Furthermore, Grad-CAM visualizations validated the clinical interpretability of the system, as heatmaps consistently indicated anatomically correct areas related to cataract pathology. This work introduces a scalable, clinically viable, and resource-efficient AI framework for cataract screening, which provides the groundwork for wider dissemination in ophthalmic diagnostics.