For the external validation cohorts, the researchers found slightly higher pooled sensitivity for machine learning models in comparison to deep learning models (85 percent vs. 81 percent).
“Most machine learning models in the included studies utilized LR (logistic regression) algorithms, often combined with clinical data to construct multimodal models, which may contribute to their higher diagnostic performance. However, the apparent superiority of traditional machine learning algorithms may also result from factors unrelated to the algorithm itself, such as differences in data splitting methods and feature selection bias between studies.” noted Han and colleagues.
Three Key Takeaways
- MRI-based AI models show strong diagnostic performance. The meta-analysis demonstrated pooled external validation sensitivity of 83 percent, specificity of 81 percent, and an AUC of 85 percent for predicting microvascular invasion (MVI) in hepatocellular carcinoma (HCC), suggesting robust performance even in external validation cohorts.
- Machine learning models may slightly outperform deep learning. Traditional machine learning models, particularly logistic regression algorithms combined with clinical data, showed slightly higher sensitivity (85 percent vs. 81 percent) compared to deep learning models, potentially due to their integration with clinical features.
- Integrating clinical data improves predictive accuracy. Models that combined radiomics with clinical factors achieved higher sensitivity (82 percent vs. 80 percent) and specificity (84 percent vs. 79 percent) than radiomic-only models, highlighting the value of multimodal approaches in enhancing prediction of MVI in HCC.
In comparison to radiomic-only models, the meta-analysis authors determined that models that combine radiomics and clinical factors offered higher sensitivity (80 percent vs. 82 percent) and specificity (79 percent vs. 84 percent).
“These findings suggest that integrating clinical features with radiomic characteristics provides a more comprehensive reflection of the tumor's biological heterogeneity, thereby enhancing the predictive accuracy for MVI. Potential mechanisms for this phenomenon include the ability of clinical features, such as AFP (alpha-fetoprotein) levels and liver function indicators, to reflect the patients' pathological and physiological complexities, complementing the imaging characteristics and improving the model's discriminative power and robustness,” explained Han and colleagues.
(Editor’s note: For related content, see “Key MRI Findings Predictive of Treatment Response for Unresectable Hepatocellular Carcinoma,” “Could an Emerging PET Tracer be a Game Changer for Detecting Hepatocellular Carcinoma?” and “Seven Takeaways from New Literature Review on Ultrasound and Hepatocellular Carcinoma.”)
In regard to limitations of the meta-analysis, the authors acknowledged that most of the reviewed studies were retrospective and conceded they only included data from the AI algorithm that had the highest sensitivity in each study. Noting that all of the reviewed studies had Chinese cohorts, the researchers said the meta-analysis findings may have limited extrapolation to broader populations.
