Multi-task deep learning based on T2-Weighted Images for predicting Muscular-Invasive Bladder Cancer

Abstract

Background: An accurate preoperative assessment of Non-Muscle-Invasive Bladder Cancer (NMIBC) and Muscle-Invasive Bladder Cancer (MIBC) in Bladder Cancer (BCa) can help the urologist make diagnostic decisions. Considering the absence of multiparametric MRI for contrast medium allergy and economic reasons, this study aims to develop a deep learning method based on T2-Weighted (T2WI) images alone for predicting NMIBC and MIBC.

Method: We propose a Multi-task BCa Muscular Invasion Prediction (MBMIP) model to discriminate MIBC from NMIBC. The three-channel-input including the original T2WI image, segmented bladder, and the region of interest can help the MBMIP model locate the bladder and pay more attention to the surrounding information of the tumor. Inception V3 is used as the feature extraction module, which uses multiple branches to extract high-level features with different degrees of abstraction. In addition, based on the idea of multi-task learning, a reconstruction block for T2WI images is also introduced to assist the backbone classification network to improve the classification performance.

Results: The entire data consist of retrospective data (390 cases), prospective data (39 cases), and multi-center data (39 cases). In the retrospective test, the accuracy, sensitivity, and specificity of the MBMIP model are 0.911, 0.889, and 0.920 respectively, while those of the prospective test are 0.923, 1.000, and 0.885. And in the muti-center test, the MBMIP model yields accuracy, sensitivity, and specificity of 0.846, 0.667, and 0.879.

Conclusion: The MBMIP model could achieve a satisfactory prediction result in discriminating between NMIBC and MIBC, which may aid urologists in preoperative decision-making for BCa patients.

Keywords: Bladder cancer; Deep learning; Image classification; Multi-task learning; Muscular invasiveness.