Multimodal Feature Learning Module for Prostate Cancer Prediction

Abstract

In worldwide, prostate cancer is one of the frequently identified cancer.  Now a days, It is very common in male with above 40 years. The accurate prediction of risk, progression and response to treatment remains challenging issues. Traditional model of predictions primarily depending on individual data such as prostate specific antigen (PSA), biopsy or medical imaging characteristics. The advances of Artificial Intelligence (AI) potentially improved cancer prediction by integrating heterogeneous healthcare information. Effective fusion of multimodal data remains a challenging problem due to the differences in data values, dimensionality and biological significance. This research proposes a Multimodal Feature Learning Module (MFLM) for the prediction of prostate cancer.  This Multimodal module integrates deep feature learning from clinical, magnetic resonance imaging (MRI) and genomic data. It employs modality specific feature extraction with an attention based fusion mechanism to learn corresponding relationships from different data modalities. The proposed framework uses deep learning model, including convolutional neural networks (CNNs), transformer based encoders and fully connected system to extract useful patterns from heterogeneous patient information.