Ovarian cancer detection and classification using attention-based CNN and deep feature extraction techniques
Ovarian cancer (OC) remains a significant global health concern, marked by high mortality rates and a lack of reliable diagnostic tools. Early detection of this disorder is crucial for improving survival rates and optimizing the healthcare system. However, it is necessary to establish an automated system that is more accurate and consistent for decision-making. In this paper, we proposed an attention-based CNN model for classifying ovarian cancer. The proposed model is evaluated on the STRAMPN dataset, which comprises 987 histopathological images. Initially, various preprocessing techniques were extensively analyzed to determine the most suitable image representation, and data augmentation schemes were applied to enhance model robustness. After that, a modified attention-based ResNet50 architecture was proposed and fine-tuned through transfer learning to extract meaningful features from the training samples. Then the SHAP method was employed for feature selection to determine the most pertinent features. Finally, we utilized an attention-based CNN model to classify the images of OC. The proposed model achieved an accuracy of 99.09%, precision of 99.17%, recall of 99.21%, F1-score of 99.06%, and an AUC of 99.85%, with a training time of 1.93 seconds, demonstrating strong computational efficiency. Experimental results demonstrate that the proposed model outperforms existing methods based on quantitative evaluation. Thus, it could assist clinicians in diagnosing OC in clinical contexts. Furthermore, the proposed model is incorporated into a web application utilizing the FastAPI framework to facilitate real-time predictions. The web application can be accessed through the following link: https://ovarian.francisrudra.com .