Deep learning for hepatocellular carcinoma segmentation in MRI: A systematic review of models, performance, and challenges.
Authors
Affiliations (4)
Affiliations (4)
- Department of Medical Physics, School of Medicine, Isfahan University of Medical Sciences, Isfahan, Iran.
- Department of Medical Physics, School of Medicine, Ahvaz Jundishapur University of Medical Sciences, Ahvaz, Khuzestan, Iran.
- Department of Medical Physics, School of Medicine, Iran University of Medical Sciences, Tehran, Iran.
- Department of Radiooncology, School of Medicine, Cancer Prevention Research Center, Seyyed Al-Shohada Hospital, Isfahan University of Medical Sciences, Isfahan, Iran.
Abstract
Hepatocellular carcinoma (HCC) is a leading cause of cancer-related mortality, necessitating accurate segmentation in magnetic resonance imaging (MRI) for diagnosis and treatment planning. Deep learning (DL) models, particularly convolutional neural networks, have shown promise in automating HCC segmentation, yet challenges like dataset limitations and MRI protocol variability persist. This systematic review evaluates DL models for HCC segmentation in MRI, focusing on model architectures, performance metrics, and implementation challenges. Following preferred reporting items for systematic reviews and meta-analyses guidelines, we searched PubMed, Scopus, Web of Science, and Cochrane Library for peer-reviewed studies using DL for HCC segmentation in MRI. Inclusion criteria required quantitative metrics (e.g., dice similarity coefficient [DSC]) and human subjects. Two reviewers conducted screening, data extraction, and quality assessment using quality assessment of diagnostic accuracy studies-2. Narrative synthesis grouped studies by architecture and MRI sequence, analyzing performance and challenges. Of 2462 records, 13 studies met criteria, predominantly using U-Net-based models (e.g., nnU-Net, UNet++). DSCs ranged from 0.61 to 0.954, with transformers and hybrid models showing adaptability. Clinical applications included diagnosis, treatment planning, and risk assessment. Challenges included small datasets (e.g., 19-602 patients), lesion heterogeneity, and MRI protocol variability, limiting generalizability. High risk of bias in patient selection was noted in 8 studies. DL models demonstrate robust HCC segmentation performance in MRI, but dataset limitations, lesion variability, and imaging inconsistencies hinder clinical adoption. Multi-center datasets, standardized protocols, and hybrid approaches integrating radiologist input are critical for advancement.