Multiparametric MRI-based prostate cancer classification using transfer learning, feature fusion, and ensemble methods.

Abstract

Prostate cancer (PCa) is a major contributor to cancer mortality in the male population globally, and precise classification of PCa into clinically significant (CS) and clinically insignificant (CiS) is essential for selecting suitable therapeutic approaches. Multiparametric magnetic resonance imaging (mpMRI) has proven to be an indispensable technique for detecting and characterising PCa lesions. This study proposes an effective framework to classify PCa using transfer learning with pre-trained VGG19 and Vision Transformer (ViT) models for feature extraction from nine mpMRI sequences of the PROSTATEx dataset, followed by feature fusion and support vector machine (SVM) classification. An ensemble approach combining multiple outputs is employed to obtain the final classification result. The proposed framework demonstrated promising results, with the ensemble output achieving an AUC score of 0.85. Key findings include the importance of incorporating higher b-values in DWI sequences and the effectiveness of combining CNN-based and transformer-based features extracted by VGG19 and ViT models. The framework has the potential to support radiologists in accurate and efficient PCa diagnosis by differentiating between CS and CiS lesions.