Collaborative Attention-Enhanced Depthwise Residual Network for High-Precision Image Classification

Abstract

Accurate differentiation between benign and malignant ovarian tumors remains a critical challenge in clinical diagnosis due to the low contrast, noise, and morphological variability of ultrasound images. To address these challenges, this paper proposes a lightweight deep learning model-CAM-DResNet-that integrates a depthwise separable convolution structure with a collaborative attention mechanism. The proposed model enhances both computational efficiency and feature discrimination by replacing standard convolutions in ResNet with depthwise convolutions and embedding a Collaborative Channel-Spatial-Pixel Attention (CCSPA) module after each network stage. This design enables the model to effectively capture both global contextual information and local lesion details, improving its sensitivity to complex tumor boundaries. The dataset comprises 860 clinical ultrasound images from Tianjin Medical University General Hospital, including 410 benign and 450 malignant cases. Experimental results show that the CAM-DResNet achieves a classification accuracy of 91.86%, outperforming classical models such as ResNet50, DenseNet, and MobileNet. Moreover, it reduces model parameters by 33% compared to ResNet50 while maintaining high specificity (88.73%) and F1 score (91.37%). The results demonstrate that the proposed architecture can robustly identify ovarian tumor characteristics and provide a reliable, computationally efficient diagnostic tool for clinical applications.