Enhanced Ovarian Cyst Classification Using a Deep Learning Model

Ovarian carcinoma is a common cause of gynecological cancer death where differentiation of subtype plays an important role in selection of treatment. This work proposes a lightweight deep learning framework for multiclass classification of epithelial ovarian carcinoma subtypes from histopathological whole-slide images. The approach employs MobileNetV2 as a feature extractor trained in TCGA and UBC-OCEAN training data to automate the classification of ovarian cancer subtypes from histopathological images. A comparative analysis was made with ResNet-50, a highcapacity model which showed that MobileNetV2 achieved improved diagnostic accuracy (89.58 %) making it suitable for use in clinical settings with limited resources. The models were fine-tuned using transfer learning techniques, incorporating data augmentation, and optimized with the AdamW optimizer and cross-entropy loss. Precision, recall, F1-score, ROC-AUC, and confusion matrices were used to assess performance, and the results showed constant subtype- level consistency. These findings highlight lightweight CNNs' potential for workflows in digital pathology. By incorporating both high-capacity and lightweight models, the system can adapt to different clinical environments, even those with limited computing resources.