An optimized multi-scale dilated attention layer for keratoconus disease classification.

Abstract

Introduction: Keratoconus (KCN) is a progressive and non-inflammatory corneal disorder characterized by thinning and conical deformation of the cornea, resulting in visual impairment. Early and accurate detection is crucial to prevent disease progression. Conventional diagnostic methods are time-consuming and depend on expert evaluation. This study introduces an advanced deep learning (DL) model aimed at automating KCN detection using corneal topography images.

Materials and methods: The proposed model, Optimized MSDALNet, integrates a Multi-Scale Dilated Attention Layer (MSDAL) to capture local and global corneal features at varying spatial resolutions. Training is optimized using Arctic Puffin Optimization (APO), a metaheuristic algorithm inspired by puffin foraging behavior. The model includes Explainable AI (XAI) capabilities using Grad-CAM for visual interpretability. Experiments were conducted using a public KCN dataset with over 1,100 labeled corneal topography images categorized into Normal, Suspect, and KCN classes. Standard pre-processing, data augmentation, and performance evaluation metrics (accuracy, precision, recall, specificity, FNR, MCC, AUC) were applied.

Results: The Optimized MSDALNet achieved superior classification performance with an accuracy of 99.5%, precision of 99.4%, and specificity of 98.4%. The proposed model outperformed existing methods such as CNN, ViT, and Swin Transformer in terms of accuracy, computational cost (1.2 GFLOPs), and inference speed (8.4 ms/image). Grad-CAM visualization confirmed the model's focus on clinically relevant corneal regions. An ablation study demonstrated the impact of each component in the proposed framework.

Conclusion: The Optimized MSDALNet combined with APO delivers an effective and interpretable solution for KCN detection. The model excels in feature extraction, computational efficiency, and clinical transparency. Limitations include dataset size and lack of multimodal inputs. Future work will focus on incorporating diverse datasets and additional patient data to enhance generalizability and diagnostic robustness.