MRI Radiomics Score And ADC Improve HCC Recurrence Prediction

Key Takeaways

• Lower ADC values and higher radiomics scores were associated with early recurrence, and their combination showed better discrimination than either metric alone.
• Larger tumor size, multiple lesions, lower ADC values, and higher radiomics scores were independently associated with shorter disease-free survival.
• The radiomics cutoff separated high-risk and low-risk groups, with marked differences in recurrence and disease-free survival within this cohort.

A retrospective single-center study included 200 patients with HCC who underwent curative-intent resection between January 2021 and January 2025, with follow-up reviewed through January 2026. Early recurrence occurred in 66 patients, representing 33.0% of the cohort. ADC was measured with a region-of-interest approach on the section showing the largest solid viable tumor component, and the predefined MRI-based radiomics score came from an institutional pipeline using ADC maps and contrast-enhanced MRI without new feature selection, model training, or retraining. Imaging review was performed by a single radiologist blinded to recurrence status and survival outcomes. The analysis focused on 2 endpoints: early recurrence and disease-free survival.

Multivariable logistic regression identified tumor size, standardized ADC, and standardized radiomics score as independent predictors of early recurrence. Effect estimates were OR 1.55 for tumor size, OR 0.29 per 1-standard-deviation increase in ADC, and OR 4.74 per 1-standard-deviation increase in radiomics score, with all P values below 0.001. ROC analysis showed AUC values of 0.707 for ADC, 0.767 for the radiomics score, 0.837 for the combined ADC–radiomics score, and 0.892 for the full multivariable model. The combined score outperformed ADC alone with P below 0.001 and radiomics alone with P = 0.008, while the full multivariable model exceeded the combined score with P = 0.001. Discrimination was better with the combined imaging score, although the full multivariable model performed best.

For disease-free survival, larger tumor size, multiple lesions, lower ADC values, and higher radiomics scores were independently associated with shorter follow-up free of recurrence. Hazard ratios were 1.28 for tumor size, 1.68 for multiple lesions, 0.52 per 1-standard-deviation increase in ADC, and 2.55 per 1-standard-deviation increase in radiomics score, with all but lesion number showing P values below 0.001. Using a Youden-derived radiomics cutoff of 0.62, 97 patients were classified as high risk and 103 as low risk. Recurrence occurred in 56 of 97 high-risk patients and 10 of 103 low-risk patients, and the high-risk group had significantly worse disease-free survival with log-rank P below 0.001. Separation between the radiomics-defined groups was substantial within this cohort.

Interpretation is shaped by several source-stated limits. The study was retrospective and single center, lacked external validation, and used a predefined radiomics score that was not redeveloped in the current cohort. The Youden-derived radiomics cutoff was also determined and tested within the same cohort, which may have led to optimistic separation of disease-free survival curves. Imaging assessment relied on a single reader without interobserver testing, and the original segmentation and feature-construction pipeline could not be fully re-evaluated from the retrospective dataset. Pathologic variables linked to recurrence, including microvascular invasion, differentiation grade, capsule invasion, and margin status, were not incorporated into the predictive models. The investigators indicated that external validation is needed before routine clinical implementation.