On evaluating CT image enhancement techniques for deep learning based 3D liver segmentation
Sidra Gul, Muhammad Salman Khan · 2022
Medical image pre-processing plays a crucial role in designing a robust computer-aided diagnostic system. The liver is difficult to segment from volumetric images because its intensity values are similar to those of nearby organs. The pre-processing step aids in the accurate feature extraction of an organ such as the liver from the given data and reduces false positive results, which reduces error in the later stages of diagnosis. In this paper, three image enhancement techniques, Hounsfield Unit (HU), contrast limited histogram equalization (CLAHE), and the combination of Hounsfield Unit and CLAHE (HU+CLAHE), have been evaluated for 3D computed tomographic (CT) volumetric images from the LiTS17 dataset to enhance image quality. We have first applied the HU and CLAHE separately as well as a pipeline of HU and then CLAHE that is first HU and then CLAHE is used to pre-process the images. After pre-processing, liver is segmented using 3D V-Net, a deep learning neural network. The 3D V-Net is trained end-to-end using the pre-processed 3D volumes obtained from the schemes described above. Dice coefficient has been used to evaluate the three pre-processing pipelines and scores achieved for CLAHE, HU and HU+CLAHE are 0.970, 0.964 and 0.971 respectively.