An inception-based deep multiparametric net to classify clinical significance MRI regions of prostate cancer

Objective.Multi-parametric magnetic resonance imaging (MP-MRI) has played an important role in prostate cancer diagnosis. Nevertheless, in the clinical routine, these sequences are principally analyzed from expert observations, which introduces an intrinsic variability in the diagnosis. Even worse, the isolated study of these MRI sequences trends to false positive detection due to other diseases that share similar radiological findings. Hence, the main objective of this study was to design, propose and validate a deep multimodal learning framework to support MRI-based prostate cancer diagnosis using cross-correlation modules that fuse MRI regions, coded from independent MRI parameter branches.Approach.This work introduces a multimodal scheme that integrates MP-MRI sequences and allows to characterize prostate lesions related to cancer disease. For doing so, potential 3D regions were extracted around expert annotations over different prostate zones. Then, a convolutional representation was obtained from each evaluated sequence, allowing a rich and hierarchical deep representation. Each convolutional branch representation was integrated following a special inception-like module. This module allows a redundant non-linear integration that preserves textural spatial lesion features and could obtain higher levels of representation.Main results.This strategy enhances micro-circulation, morphological, and cellular density features, which thereafter are integrated according to an inception late fusion strategy, leading to a better differentiation of prostate cancer lesions. The proposed strategy achieved a ROC-AUC of 0.82 over the PROSTATEx dataset by fusing regions ofKtransand apparent diffusion coefficient (ADC) maps coded from DWI-MRI.Significance.This study conducted an evaluation about how MP-MRI parameters can be fused, through a deep learning representation, exploiting spatial correlations among multiple lesion observations. The strategy, from a multimodal representation, learns branches representations to exploit radio-logical findings from ADC andKtrans. Besides, the proposed strategy is very compact (151 630 trainable parameters). Hence, the methodology is very fast in training (3 s for an epoch of 320 samples), being potentially applicable in clinical scenarios.