Accurate segmentation of bladder wall and tumor regions in MRI using stacked dilated U-Net with focal loss

Abstract

Automatic and accurate segmentation of bladder walls and tumors in magnetic resonance imaging (MRI) is a challenging task, due to significant bladder shape variations, strong intensity inhomogeneity in urine and very high variability across tumors appearance. To tackle such issues, we propose to leverage the representation capacity of an improved U-Net networks using stacked dilated convolutions. The proposed structure includes stacked dilated convolutions to increase the receptive field without incurring gridding artifacts. In addition, we embed stacked dilated convolution network into the U-Net architecture, thus enabling extracting multi-scale features for segmentation of multi structures with different shapes and scales. Finally, we apply a focal loss function to make all classes contribute equally to the loss function in our model. Evaluations on T2-weighted MRI show the proposed model achieves a higher level of accuracy than state-of-the-art methods, with a mean Dice similarity coefficient of 0.95, 0.81 and 0.66 for inner wall, outer wall and tumor region segmentation, respectively. These results demonstrate a strong agreement with reference standards and a high performance gain compared with existing methods.