Society: AASLD

Background: The LI-RADS (Liver Reporting and Data Systems) was created to standardize reporting of CT/MR imaging for liver masses in cirrhosis patients at risk of hepatocellular carcinoma (HCC). LI-RADS (LR) 3 and 4 observations are indeterminate and have a variable natural history, with 35-65% becoming HCC and others remaining stable or regressing over time. We hypothesize that by using artificial intelligence and radiomics, we can extract imaging features to better predict presence of HCC in patients with LR3/4 lesions.
Method: Leveraging the analytic morphomics image processing platform, we used a random sampling strategy to extract radiomics features on automatically registered images from different contrast phases. Samples were randomly extracted from the tumor and peritumoral areas. Random virtual nodules were computationally generated in nontumoral areas and in control patients with cirrhosis. Chart review was performed for extraction of clinical data and outcomes. Multi-modal prediction models were developed using gradient boosting tree framework with automated hyperparameter selection.
Results: We used a retrospective cohort of 391 patients with cirrhosis and a pre-treatment multi-phase CT scan (219 from University of Michigan and 172 from University of Texas Southwestern) to train and internally validate models. Mean age was 58.6 years, 71.7% were male, and cohort was diverse regarding race/ethnicity (57.3% white, 17.5% Hispanic, 17.0% Black, 6.2% other) and liver disease etiology (56.3% viral, 16.7% alcohol, 15.2% NAFLD, 11.8% other). 252 (64.5%) patients were diagnosed with HCC within 1 year of the index scan, 1 segmented nodule and 1 virtual nodule was used per patient, comprising 21 LR3, 48 LR4, and 183 LR5 nodules. For the 139 controls, 2 virtual nodule per patient was utilized with the exception of 1 LR3 nodule which did not become HCC. Model performance was validated using k-fold cross validation. We found that models that included radiomics outperformed AFP alone in all cases (AUROC 0.89 – 0.91 vs. 0.80) (Table). The best multi-modal models which included radiomics on index imaging with AFP at time of index imaging performed similarly in LR3 (AUROC 0.91) and LR4 (AUROC 0.90) observations compared to those with LR 5 (AUROC 0.92) in predicting HCC (Figure).
Conclusion: Artificial intelligence and radiomics are promising tools to accurately predict risk of HCC in indeterminate liver nodules.
Funding Sources: NIH U01CA230669; VA I01HX002548; U01 CA230694

<b>TABLE: </b>Performance of Different Models to Predict HCC Using Radiomics and Clinical Data

TABLE: Performance of Different Models to Predict HCC Using Radiomics and Clinical Data

<b>Figure</b>. AUROC of the Full Model in Prediction of HCC in LR3,LR4,vs LR 5 observations

Figure. AUROC of the Full Model in Prediction of HCC in LR3,LR4,vs LR 5 observations

Sven Holcombe

Appearances