Adaptable volumetric liver segmentation model for CT images using region-based features and convolutional neural network

Liver plays an important role in metabolic processes, therefore fast diagnosis and potential surgical planning is essential in case of any disease. The automatic liver segmentation approach has been studied during the past years and different segmentation techniques have been proposed, but this task remains a challenge and improvements are still required to further increase segmentation accuracy. In this work, an automatic, deep learning based approach is introduced, which is adaptable and it is able to handle smaller databases, including heterogeneous data. The method starts with a preprocessing to highlight the liver area using probability density function based estimation and supervoxel segmentation. Then, a modification of the 3D U-Net is introduced, which is called 3D RP-UNet and applies the ResPath in the 3D network. Finally, with liver-heart separation and morphological steps, the segmentation results are further refined. Segmentation results on three public databases showed that the proposed method performs robustly and achieves good segmentation performance compared to other state-of-the-art approaches in the majority of the evaluation metrics.