Novel uncertainty driven boundary refined convolution neural network for uneven medical image segmentation

Abstract

Organ segmentation has become a fundamental challenge for computer-aided intervention, diagnosis, radiation therapy, and critical robotic surgery. Automatic organ segmentation from medical images is easy for large organs with regular shape. However, it becomes challenging task due to the inconsistent shape and size of different organs. Additionally, low contrast at the organ borders, resulting from analogous tissue types, impairs the network’s capacity to accurately delineate organ contours. In this thesis, we propose an end-to-end uncertainty-driven boundary-refined segmentation network (UDBRNet) for segmenting organs from computed tomography (CT) images. This network consists of three modules. Firstly, an encoder-decoder based segmentation module generates a main and two auxiliary segmentation masks using multi-line decoders, and then, uncertainty is assessed using variations in the levels of agreement and disagreement among the masks in uncertainty determination module. Finally, both the main segmentation mask and the uncertainty information are sent to a boundary refinement module, which refines organs’ boundary residuals. The uncertainty information helps the UDBRNet’s boundary refinement module to improve the low contrast and inconsistently shaped organs’ edge refinement. Our proposed segmentation network demonstrates remarkable performance, with dice accuracies of 0.80, 0.95, 0.92, and 0.94 for Esophagus, Heart, Trachea, and Aorta respectively on the SegThor dataset. It also consistently shows superior performance, with dice accuracies of 0.71, 0.89, 0.85, 0.97, and 0.97 for Esophagus, Spinal Cord, Heart, Left-Lung, and Right-Lung respectively on the LCTSC dataset. We compare the UDBRNet performance with popular eight existing state-of-the-art segmentation methods: UNet, attentionUNet, BASNet, FC-denseNet, R2UNet, UNet++, TransUNet, and DS-TransUNet using two publicly available datasets, SegTHor and LCTSC. Results show that UDBRNet outperforms those existing methods. UDBRNet presents a promising network for more precise organ segmentation, particularly in challenging, uncertain conditions.