Semantic Organ Segmentation using Regression Solutions (Loss Functions) based on Organ Distance Map
Hossein Arabi, Habib Zaidi · 2022
Convolutional neural networks have been widely proposed/developed to conduct efficient and accurate organ segmentation, wherein a variety of loss functions have been considered to achieve optimal organ detection from the input data. Despite extraordinary performance of neural networks, noticeable errors located at the boundary/proximity of the target organs are observed as well as outliers in the form of island of voxels or voids/holes within the solid/uniform part of the target. To address these issues, we set out to investigate an alternative semantic segmentation solution based on the distance map using the regression solutions. In the proposed framework the distance maps obtained from the target masks are directly considered as the target within the training of the model, wherein regression networks and loss functions are used for the development of the model. Since the distance maps are continuous-valued data (as opposed to binary masks) any regression solutions could be employed to achieve the segmentation of the target organ/lesion. To implement the regression-based segmentation framework, a residual neural network was trained using the conventional segmentation framework and a hybrid Jac-card/cross-entropy loss function and the distance map framework using an L2 loss function. These models were tested for kidney segmentation using the decathlon dataset. The regression-based approach outperformed the conventional deep learning-based segmentation method with statistically significant differences (p-values <0.05). In addition to the overall superior performance of the regression-based approach, this method led to no outliers in the form of islands of voxels or holes inside the target structure. However, the conventional segmentation approach resulted in 5 outliers out of 30 subjects in the test dataset. Moreover, owing to the fact that the segmentation problem is converted to a regression task, we employed an L2 loss function since this function would perform better if the training dataset were not noisy compared to the L1 loss function. We intentionally added some noise and wrong samples to the training dataset and repeated the training. In this case, the L1 loss function led to better results than the L2 loss function. This would be regarded as an advantage of the proposed framework that the regression loss function with specific properties could be employed for the segmentation tasks. Overall, this study demonstrated the superior performance of the proposed regression-based solution over the conventional deep learning segmentation approach.