Application of deep learning as a noninvasive tool to differentiate muscle-invasive bladder cancer and non-muscle-invasive bladder cancer with CT

Objective: To construct a deep-learning convolution neural network (DL-CNN) system for the differentiation of muscle-invasive bladder cancer (MIBC) and non-muscle-invasive bladder cancer (NMIBC) on contrast-enhanced computed tomography (CT) images in patients with bladder cancer. Materials and methods: A total of 1200 cross-sectional CT images were obtained from 369 patients with bladder cancer receiving radical cystectomy from January 2015 to June 2018, including 249 non-muscle-invasive bladder cancer (NMIBC) series and 120 muscle-invasive bladder cancer (MIBC) series. All eligible images were distributed randomly into the training, validation, and testing cohorts with ratios of 70 %, 15 %, and 15 %, respectively. We developed one small DL-CNN containing four convolutional and max pooling layers and eight DL-CNNs with pretrained bases from the ImageNet dataset to differentiate NMIBC from MIBC. The intermediate activations were applied on the test dataset to visualize how successive DL-CNN layers transform their input. Results: The area under the receiver operating characteristic curve (AUROC) of the validation and testing datasets for the small DL-CNN was 0.946 and 0.998, respectively. The AUROCs of eight deep learning algorithms with pretrained bases ranged from 0.762 to 0.997 in the testing dataset. The VGG16 model had the largest AUROC of 0.997 among the eight algorithms with a sensitivity and specificity of 0.889 and 0.989. The independent features encoded by the small DL-CNN filters were displayed as assemblies of individual channels. Conclusion: Based on contrast-enhanced CT images, our DL-CNN system could successfully classify NMIBC and MIBC with favorable AUROC in patients with bladder cancer. The application of our system in early stage might assist the pathological examination for the improvement of diagnostic accuracy.