Automated liver and liver tumor segmentation based on deep learning for 3D computed tomography of patients with colorectal liver metastasis

Abstract

This project focuses on the development of a reliable and automated 3D segmentation pipeline for liver and colorectal liver metastases (CRLM) using computed tomography (CT) scans. The study combines a detailed literature review on CRLM and deep learning methods with an extensive experimental process implemented in the MONAI framework. Multiple architectures, strategies and parameters were explored, with SegResNet emerging as the most effective model for tumor segmentation. The proposed two-stage pipeline first segments the liver and then uses the liver mask to guide the tumor segmentation task. Particular attention was given to preprocessing and data augmentation to address issues such as class imbalance, depth variation, and heterogeneous tumor appearances. Experiments were conducted on an optimized dataset, excluding patients with very few tumor slices, using Dice similarity coefficient (DSC), recall, precision, and surface distance metrics for evaluation. The final liver model achieved state-of-the-art performance with a Dice score of 0.968, while the tumor segmentation model reached 0.674 Dice, a competitive result given the difficulty of the task and the limited data and hardware resources. Overall, the project demonstrates the potential of deep learning and 3D medical imaging for the accurate segmentation of CRLM, providing a solid foundation for future research using larger datasets and more specialized models.