Gastroenterology & Endoscopy · Published 2025-08-23 · DOI 10.1016/j.gande.2025.08.003
Accurate extraction and segmentation of the liver from medical images remain challenging due to the organ's irregular shape, similar intensity values to adjacent structures, variable contrast levels, and inherent noise in CT images. Addressing these issues is crucial for effective liver analysis and early cancer detection. This study investigates the impact of various pre-processing techniques—including windowing, normalization, histogram equalization, median filtering, and block filtering—applied to 3D CT volumes from the publicly available Liver Tumor Segmentation Benchmark (LiTS) dataset. These pre-processing methods aim to enhance image quality and facilitate more accurate liver segmentation using deep learning models. The results demonstrate that the optimized pre-processing pipeline significantly improves segmentation performance, achieving a Dice similarity coefficient of 77.15 % and an overall accuracy of 93 %. Moreover, the application of these techniques contributed to increasing the mean positive predictive value (PPV) to 0.94, indicating more reliable identification of liver structures. Future work will focus on implementing the UNET architecture for precise segmentation, with the goal of achieving higher accuracy and aiding early diagnosis of liver cancer. The findings underscore the importance of pre-processing in overcoming imaging challenges and improving deep learning-based detection systems, ultimately contributing to timely intervention and better patient outcomes.
Abstract from DOAJ. Public domain (CC0 1.0).
Read the article at the publisher →