Multi-path neural network based on mp-MRI for predicting muscle-invasive bladder cancer

Abstract

Purpose

To assist doctors in clinical diagnosis, we propose a multipath deep learning (MP-DL) model to distinguish between muscle-invasive bladder cancer (MIBC) and non-muscle-invasive bladder cancer (NMIBC) using multiparametric magnetic resonance imaging (mp-MRI) and synthesized samples.

Methods

The proposed MP-DL model integrates T2-weighted image (T2WI), diffusion-weighted imaging (DWI), and dynamic contrast-enhanced imaging (DCE) branches, combining high-level features from these sequences. Multichannel inputs, including the original image, segmented bladder, and region of interest from the T2WI branch, enhance the focus on the tumor region. InceptionV3 served as the model backbone for feature extraction with a multitask framework in the DWI and DCE branches. Synthesized samples were generated to supplement missing DWI and DCE sequences, enlarging the sample size and boosting model performance. Evaluation through five-fold cross-validation and internal-external testing demonstrated the model's robustness and generalization capabilities.

Results

The dataset consisted of 401 cases (287 NMIBC and 114 MIBC cases), which included a training set of 313 cases (containing partly synthesized DWI and DCE samples), validation of 26 cases, internal testing of 34 cases, and external testing of 28 cases. In the internal testing, the model achieved area under curve, accuracy, sensitivity, specificity, and F1 scores of 0.914, 0.835, 0.880, 0.817, and 0.761, respectively. In the external testing, the results were 0.821, 0.814, 0.730, 0.845, and 0.619, respectively. The model performance surpasses that of medical professionals in internal testing while outperforming urologists and approaching senior radiologists in external testing.

Conclusions

The MP-DL model, incorporating synthesized samples, shows promise in preoperative MIBC prediction, potentially aiding primary urologists and radiologists.