Abstract
Precise segmentation of bladder walls and tumor regions is an essential steptowards non-invasive identification of tumor stage and grade, which is criticalfor treatment decision and prognosis of patients with bladder cancer (BC).However, the automatic delineation of bladder walls and tumor in magneticresonance images (MRI) is a challenging task, due to important bladder shapevariations, strong intensity inhomogeneity in urine and very high variabilityacross population, particularly on tumors appearance. To tackle these issues,we propose to use a deep fully convolutional neural network. The proposednetwork includes dilated convolutions to increase the receptive field withoutincurring extra cost nor degrading its performance. Furthermore, we introduceprogressive dilations in each convolutional block, thereby enabling extensivereceptive fields without the need for large dilation rates. The proposednetwork is evaluated on 3.0T T2-weighted MRI scans from 60 pathologicallyconfirmed patients with BC. Experiments shows the proposed model to achievehigh accuracy, with a mean Dice similarity coefficient of 0.98, 0.84 and 0.69for inner wall, outer wall and tumor region, respectively. These resultsrepresent a very good agreement with reference contours and an increase inperformance compared to existing methods. In addition, inference times are lessthan a second for a whole 3D volume, which is between 2-3 orders of magnitudefaster than related state-of-the-art methods for this application. We showedthat a CNN can yield precise segmentation of bladder walls and tumors inbladder cancer patients on MRI. The whole segmentation process isfully-automatic and yields results in very good agreement with the referencestandard, demonstrating the viability of deep learning models for the automaticmulti-region segmentation of bladder cancer MRI images.