GCAP-Net: Improving MRI-Based Ternary Classification of Ovarian Tumors with Deep Learning

Accurate classification of ovarian tumors is crucial for guiding clinical treatment decisions and assessing patient prognosis. However, most existing studies focus on binary classification tasks, distinguishing either benign from malignant or malignant from borderline tumors. To address this challenge, we propose a novel network architecture, GCAP-Net, designed to classify ovarian tumors into benign, borderline, and malignant categories. This model introduces an Adaptive Graph Channel Attention module to enhance focus on essential channel features. It also replaces traditional average pooling with a Dual-Pooling Scorer module, which integrates local and global feature information to improve discrimination of ovarian tumor types. Experimental results show that GCAP-Net achieved an accuracy of 70.77%, an area under the curve (AUC) of 0.8128, and a specificity of 85.04% on the test set. These results demonstrate strong performance in ovarian tumor classification tasks.