Automated detection of prostate cancer from multiparametric MRI using deep convolutional neural networks

Abstract

Prostate cancer (PCa) is the second most common cause of cancer-related deaths among males over 50 years. As the most advanced non-invasive tool for characterization of PCa, Multiparametric Magnetic Resonance Imaging (mpMRI) compounds dynamic contrast-enhanced (DCE), diffusion-weighted(DWI), T1- (T1W) and T2-weighted(T2W) imaging modalities to give precise and in-depth tumour information. In practice, lesions are assigned scores (1–5) which signify prostate cancer severity based on clinical significance using the Prostate Imaging-Reporting and Data System (PI-RADS); however, the interpretation of mpMRI using this approach is susceptible to inconsistencies caused by inter-reader and intra-reader variability. These challenges are particularly pronounced in low- and medium-income areas where the availability of specialized radiologists is limited. This project aimed at automating the detection of prostate cancer from multiparametric MRI using deep convolutional neural networks (DCNNs). The dataset used consisted of T2W, ADC and DWI image sequences categorised into groups based on absence or presence of lesions, as well as their clinical significance. The metrics for evaluating the classification and segmentation models developed were specificity, sensitivity, Dice Similarity Coefficient(DSC) and Intersection Over Union(IOU). The models were trained on data of 143 patients, and the selection of the best models based on test results obtained using a virgin set of 20 patients. The best slice and lesion classifiers achieved sensitivity of 85% and 100% with T2W test images, and 77.8% sensitivity on DWI test images containing lesions in the TZ and other zones of the gland respectively. The best UNet models produced masks with DSC of 0.77 and 0.737 for central and transition zone segmentation of T2W test images respectively. The best models were deployed on a web based platform using Flask framework. The system can be used for early detection of prostate cancer from mpMRI.