Integrative Multi-Stage Deep Learning Framework for Ovarian Tumor Ultrasound Classification with Explainability and Confidence Estimation

Shtwai Alsubai^1*Ahmad Almadhor²Abdullah Al Hejaili³

• ¹Prince Sattam Bin Abdulaziz University, Al-Kharj, Saudi Arabia
• ²Jouf University, Sakaka, Saudi Arabia
• ³University of Tabuk, Tabuk, Saudi Arabia

Ovarian cancer is a major diagnostic problem because it is asymptomatic at its early stages and requires subjective interpretation of an ultrasound. In this research, we present the EfficientOvaNet framework, a deep learning based model to classify ovarian tumors using ultrasound images, which was trained on the Multi-Modality Ovarian Tumor Ultrasound (MMOTU) dataset. This framework employs a two-branch EfficientNet-B3 model to combine Region-of-Interest (ROI) features with global contextual information, along with sophisticated preprocessing, data augmentation, and class-imbalance control via weighted Focal Loss. The five-fold cross-validation produces a mean accuracy of 91.9%, F1-score of 91.9% and AUC of 0.98, which indicates better performance than the baseline models. Explainable methods, such as Grad-CAM, Monte Carlo Dropout uncertainty estimation, and t-distributed Stochastic Neighbor Embedding (t-SNE) based feature visualization, ensure clinical interpretability and credibility. EfficientOvaNet can improve timely intervention and individualized treatment by increasing diagnostic accuracy and reducing subjectivity, potentially improving survival rates in the management of ovarian cancer.