Deep DPA-MSFNet: A Dual-path Attention and Multi-scale Fusion Network for Robust Eye Tumor Classification

Abstract

—The rapid ascent of deep learning in medical image analysis has been important in the remarkable advancements made in automated tumor identification. This study presents a novel convolutional neural network architecture for efficient tumor classification by integrating multi-scale fusion techniques with dual-path attention. The model employs sophisticated preprocessing techniques including Anisotropic Diffusion Filtering (ADF), Contrast Limited Adaptive Histogram Equalization (CLAHE), and SOBEL on 64×64×3 RGB input images that have already been preprocessed to enhance contrast and edge features. The Dual-Path Attention Block integrates Squeeze-and-Excitation (SE) and Convolutional Block Attention Module (CBAM) attention methods to document spatial and channel interdependence, making it a crucial component of the design. Multi-Scale Fusion Blocks enhance feature diversity and effectively capture local and global context by using parallel convolutional paths (1×1, 3×3 dilated, and 5×5). After three iterations of the design, the final prediction is made using a global average pooling layer, a lightweight classification head, and a softmax output layer. The model’s excellent accuracy and robustness in distinguishing between tumor and non-tumor regions, as shown by experimental results, making it a great fit for resource-constrained medical devices that need to be deployed in real-time.