Automatic mpMRI-based Prostate Lesions Assessment with Unsupervised Domain Adaptation

Multiparametric magnetic resonance imaging (mpMRI) has emerged as a valuable diagnostic tool in prostate lesion assessment, for which many reports conclude the advantageous use of convolutional neural networks (CNNs) in prostate lesion detection and classification (PLDC). However, the network training inevitably involves prostate magnetic resonance (MR) images from multiple sites/cohorts. There always exists variation in scanning protocol among the cohorts, inducing significant changes in data distribution between source and target domains. This challenge has greatly limited clinical adoption on a large scale. Recent domain adaptation (DA) models can alleviate the inherent cross-site heterogeneity. Some could leverage cross-domain knowledge transfer using whole-slide images, without prior knowledge of lesion regions. In this paper, we propose a coarse mask-guided deep domain adaptation network (CMD²A-Net) in order to develop a fully automated framework for PLDC using multi-cohort images. No category or mask label is required from the target domain. A coarse segmentation module is trained to cover the possible lesion-related regions, so that attention maps can be generated to dedicate the local feature extraction of lesions within those regions. As a result, the features of both prostate lesion and region can be fused to align the robust features between the source and target domains. Experiments have been performed on 512 mpMRI sets from datasets of PROSTATEx (with 330 sets) and two cohorts, A (with 74 sets) and B (with 108 sets). Using the ensemble learning, our CMD²A-Net accomplishes an AUC of 0.921 in cohort A and 0.913 in cohort B, demonstrating its transferability from a large-scale public dataset PROSTATEx to our small-scale target domains. Our results and ablation study also support the CMD²A-Net’s effectiveness in lesion classification between benign or malignant, compared to the state-of-the-art models Corresponding author(s) Email: kwokkw@hku.hk (K.W. Kwok), varv@hku.hk (V. Vardhanabhuti)