Selective feature-based ovarian cancer prediction using MobileNet and explainable AI to manage women healthcare

Polsinelli, Matteo

DOI: https://doi.org/10.1007/s11042-024-19286-6

IF: 2.577

2024-05-14

Multimedia Tools and Applications

Abstract:Ovarian cancer is a global health concern due to the unavailability of an effective screening strategy and is often diagnosed at a late stage with approximately 70% of the case which reduces the survival chances of patients. Initial diagnosis is challenging due to inconspicuous symptoms in its initial stages which complicate its timely diagnosis. Regular screenings, such as pelvic exams, ultrasounds, and blood tests targeting specific biomarkers can be helpful for early diagnosis. In addition, the use of machine learning models can help automate this process thereby assisting medical experts for accurate diagnosis. This study aims at timely and accurate detection of ovarian cancer using a transfer learning approach that uses the MobileNet model. Moreover, the Chi-square technique is used to extract the most impactful features for better accuracy. For experiments, the Soochow University ovarian cancer dataset is employed. Extensive experiments are performed using all features from the dataset, as well as, using Chi-square-based selective features. The best accuracy of 90.88% is achieved with the Xception model when all features are used. Experimental results show a substantial increase when selective features are used indicating a 98.49% accuracy using the MobileNet model with 20 most important features. In addition, precision, recall, and F1 scores of 99.13%, 99.27%, and 99.20% are obtained showing the model's robustness and generalization. This study also employs Shapley additive explanations to explain the importance of various features toward the model's output thereby providing transparency for the model's decision-making process.

computer science, information systems, theory & methods,engineering, electrical & electronic, software engineering