Integrative multi-stage deep learning framework for ovarian tumor ultrasound classification with explainability and confidence estimation

Introduction: Ovarian cancer is a major diagnostic problem because it is asymptomatic in its early stages and requires subjective interpretation of ultrasound images. Methods: This research presents the EfficientOvaNet framework, a deep learning-based model for classifying ovarian tumors using ultrasound images, trained on the Multi-Modality Ovarian Tumor Ultrasound (MMOTU) dataset. The framework employs a two-branch EfficientNet-B3 architecture that combines Region-of-Interest (ROI) features with global contextual information. Sophisticated preprocessing, data augmentation, and class-imbalance control using weighted Focal Loss are applied. Five-fold cross-validation is used for performance evaluation. Explainable methods, including Grad-CAM, Monte Carlo Dropout uncertainty estimation, and t-distributed Stochastic Neighbor Embedding (t-SNE)-based feature visualization, are incorporated to ensure interpretability. Results: The five-fold cross-validation yields a mean accuracy of 91.9%, an F1-score of 91.9%, and an AUC of 0.98, indicating better performance than baseline models. Discussion: EfficientOvaNet increases diagnostic accuracy and reduces subjectivity in ultrasound-based ovarian tumor classification. By improving interpretability and credibility, the framework has the potential to support timely intervention and individualized treatment, which may improve survival rates in the management of ovarian cancer.