Attention-Enhanced ResNet۵۰ for Robust Multi-Class Classification of Intraocular Tumors in Ultra-Widefield Fundus Images

Accurate multi-class classification of intraocular tumors remains a challenging task due to subtle inter-class variations, limited annotated datasets, and the complexity of ultra-widefield (UWF) fundus imaging. This study proposes an attention-enhanced deep learning framework based on ResNet-۵۰ for robust classification of six categories of intraocular conditions using the UWF Fundus Images of Intraocular Tumors dataset (۲,۰۳۱ images). Transfer learning was employed to leverage pretrained hierarchical representations, and systematic ablation experiments were conducted to evaluate the impact of input resolution, Squeeze-and-Excitation (SE) attention placement, and layer freezing strategies. Results demonstrated that an input resolution of ۷۶۸×۷۶۸ pixels provided an optimal balance between preserving fine pathological details and computational efficiency. Selective integration of SE attention in the deepest residual stage significantly improved generalization performance (Test Accuracy = ۰.۹۸۵, F۱-score = ۰.۹۸۶, Cohen's Kappa = ۰.۹۷۲), outperforming the baseline ResNet-۵۰. In contrast, distributing attention across multiple stages increased model complexity without consistent gains in robustness. Furthermore, strategic freezing of early residual layers functioned as an effective regularization mechanism, enhancing generalization under limited data conditions. The findings indicate that careful optimization of attention depth and parameter adaptation is critical for achieving stable and clinically reliable multi-class intraocular tumor classification.

کلیدواژه ها:

Artificial intelligence ، Deep learning ، Tumor Classification ، Intraocular Tumors ، ResNet۵۰