A multi-cancer detection framework using deep learning and hybrid machine learning approaches

Abstract

The diagnostic solutions offered by the present artificial intelligence (AI) solutions suffer from non-generalizability and heavy reliance on complex models. In an attempt to solve these issues, we propose a lightweight yet versatile method consisting of a combination of ResNet50 transfer learning and hybrid machine learning. Image features are extracted using dermoscopy, magnetic resonance imaging (MRI), and histopathological images. These are subjected to principal component analysis (PCA) dimensionality reduction followed by classification using support vector machine (SVM), random forest (RF), logistic regression (LR), and XGBoost algorithms. This segregation of the two processes improves efficiency. The hybrid approach using ResNet50 + LR yielded an accuracy of 91.01% in the case of breast cancer detection compared to 86.26% of a baseline convolutional neural network (CNN). Also, ResNet50 gave an accuracy of 96.61% in diagnosing skin cancer. Custom CNN provided an accuracy of 99.42% for lung cancer and 96.33% for brain tumor detection.