AI driven ovarian cancer early detection by subtype categorization and aberrant instances identification

Abstract

Among cancers of the female reproductive system, ovarian cancer stands as one of the most deadly types because of its ability to vary between patients and its late discovery by medical professionals. The correct identification of ovarian cancer subtypes serves both therapeutic purposes and outcome enhancement needs in medical treatment. This study presents an automated, robust classification system based on advanced machine learning architectures to overcome the challenges of subtype classification. Research on ovarian cancer subtypes still has not progressed until this work since class imbalance and uncommon subtypes were often neglected. Previously, many studies functioned under basic models and single-source data without which accuracy and generalization are limited. Specifically, performance evaluation overemphasized only accuracy while overrunning important metrics like precision and AUC. Data augmentation was underused, and early detection was often neglected. This study fills the voids to enhance diagnostic capabilities in ovarian cancer. Several deep learning architectures are considered for improving the performance of classification in this work, including pre-trained architectures like VGG16 and ResNet50, a hybrid model by taking the strengths of both, CNN, and Vision Transformers. Our hybrid model performed the best with an accuracy of 92.8%, F1-score of 0.927, recall of 92.8%, and precision of 93.5%, which showed that the integration of complementary feature extraction capabilities was e↵ective. Other advanced data preprocessing techniques, like resizing, normalization, and augmentation, are also employed in this study to improve model generalization and handle class imbalance. This work provides a very efficient yet scalable classification framework in the field of medical imaging and diagnostics of cancer. The results have pointed out the importance of hybrid architectures and pre-trained models toward superior performance and this system has great potential to be integrated into a clinical workflow; it provides a useful tool to support pathologists and oncologists in the diagnosis of ovarian cancer.