Deep learning and machine learning in image-based hepatocellular carcinoma detection: a systematic review and
Jingchao Wang1, Jingjie Bai1, Yanyang Wang1
1Department of ultrasound, Hebei Medical University Third Hospital, Shijiazhuang, China.
Background:
Early detection and management are essential in hepatocellular carcinoma (HCC) care. Artificial intelligence (AI) is enhancing the precision and efficiency of diagnostic imaging for liver tumor assessment.
Objective:
This study aimed to characterize the diagnostic performance of AI algorithms for detecting HCC via systematic review and meta-analysis.
Methods:
A systematic literature search was conducted in PubMed, Embase, Cochrane, Web of Science, and IEEE Xplore Digital Library in November 2024 to identify studies evaluating AI algorithms for detecting HCC. Diagnostic performance data from the included studies were pooled via meta-analysis and compared against human clinicians (11 studies). Subgroup analyses were conducted based on classification task (binary classification tasks (BCTs) vs. multi-class classification tasks (MCCTs)), validation type, algorithmic approach (deep learning (DL) vs. machine learning (ML)), and imaging modality (ultrasound (US), computed tomography (CT), or magnetic resonance imaging (MRI)).
Results:
Forty-one studies evaluating AI-based diagnostic models were included. The pooled sensitivity (SEN), specificity (SPC), and area under the curve (AUC) were 87% (95% CI 84-90%), 91% (95% CI 89-93%), and 0.95 (95% CI 0.92-0.96). In comparison, 11 studies assessing human clinicians yielded a pooled SEN, SPC, and AUC of 78% (95% CI 73-83%), 91% (95% CI 86-94%), and 0.91 (95% CI 0.88-0.93). These findings indicated that AI algorithms were related to high diagnostic accuracy and exhibited numerically higher performance metrics than human clinicians. Among the encompassed studies, 31 addressed BCTs and ten addressed MCCTs. Subgroup analyses across validation type, algorithmic approach, and imaging modality consistently demonstrated high diagnostic performance for AI. However, substantial heterogeneity was observed overall and within subgroups, with no clear source identified.
Conclusions:
AI algorithms are associated with high diagnostic accuracy for image-based detection of HCC, potentially surpassing human performance. However, additional prospective investigations with rigorous external validation and larger sample sizes are still necessary to detect HCC in medical imaging (MI).