Improving risk stratification and detection of early HCC using ultrasound-based deep learning models

Abstract

Background & aims: Hepatocellular carcinoma (HCC) surveillance programs are suboptimal. We aimed to design an ultrasound-based deep learning model for HCC risk stratification (STARHE-RISK) and early-stage HCC detection (STARHE-DETECT) in patients with compensated advanced chronic liver disease (cACLD).

Methods: This prospective multicentric study included 403 adult patients with cACLD of all causes enrolled in a surveillance program for at least 6 months without prior history of HCC. STARHE-RISK was trained on ultrasound cine clips of the non-tumoral liver parenchyma using two classes: cases (n = 152 patients with early-stage HCC; 137/152 [82%] male; median age 63 years) and controls (n = 170 patients without HCC at inclusion and during a subsequent 1-year follow-up; 120/170 [71%] male; median age 69 years). STARHE-DETECT was trained on tumour ultrasound cine clips. The training/validation and testing sets were stratified according to potential confounders, and 50 patients who were balanced in both groups were allocated to the independent testing set based on sample size calculation. Statistical analysis included classification and detection metrics.

Results: STARHE-RISK achieved good prediction performances in the testing set with a 0.72 accuracy (95% CI 0.57-0.84) and an odds ratio of 6.6 (95% CI 1.9-22.7; p = 0.003). The combination of STARHE-RISK and the FASTRAK score, a multi-aetiology HCC risk stratification score, achieved a higher specificity (0.86 [95% CI 0.65-0.97]) and odds ratio (8.9 [95% CI 2.1-38.3; p = 0.004]) for predicting a patient at high risk of HCC development. STARHE-DETECT achieved a 0.67 mAP10, a 0.68 sensitivity (95% CI 0.47-0.85), and a 0.82 specificity (95% CI 0.69-0.91) for detecting early-stage HCC.

Conclusions: STARHE-RISK and STARHE-DETECT achieved robust performances for HCC risk stratification and early-stage HCC detection, respectively. They could become valuable surveillance tools and pave the way for a risk-based personalised surveillance program.

Impact and implications: STARHE-RISK is a reliable ultrasound-based deep learning model for hepatocellular carcinoma (HCC) risk stratification in patients with compensated advanced chronic liver disease and can be associated with complementary scores integrating clinical and blood parameters. STARHE-DETECT could become a complementary tool to visual assessment for radiologists and sonographers in HCC surveillance. Both models are based on simple and easy-to-perform ultrasound cine clip acquisitions. This study paves the way for a risk-based personalised surveillance program that will not ultimately rely on a single test but rather on a combination of approaches mixing clinical, biological, and radiological data.

Clinical trials registration: The study is registered at ClinicalTrials.gov (NCT04802954).

Keywords: Deep learning; Hepatocellular carcinoma; Prediction; Risk stratification; Ultrasound.

Graphical abstract

Fig. 1

Flow chart of the study.

Fig. 2

Performance of STARHE-RISK and STARHE-DETECT models in the testing set. (A) Calibration curve and (B) ROC curve of the STARHE-RISK model and FASTRAK score for HCC risk stratification in the testing set. (C) Mean Average Precision (Intersection over Union of 10%, 50%, 75%) with precision-recall curve obtained by plotting the STARHE-DETECT model's precision and recall values as a function of the model's confidence score threshold.

Fig. 3

Liver parenchyma echotexture patterns and HCC risk stratification performed by STARHE-RISK. (A) Increased homogeneous echostructure in a 55-year-old man with alcohol-related cirrhosis correctly predicted at low-risk (77% confidence) compared to (B) a macronodular echostructure (arrow) in a 42-year-old man with cured hepatitis B cirrhosis and BCLC 0 hepatocellular carcinoma (not shown on the image) correctly predicted at high-risk (on this view of the non-tumoral liver parenchyma).

Fig. 4

Correctly detected HCC nodule and false positive prediction by STARHE-DETECT model in a 74-year-old patient with metabolic dysfunction-associated steatotic liver disease and 17 mm isoechoic hepatocellular carcinoma. The HCC nodule in Figure 4A was correctly detected with a confidence level of 79.4%. Figure 4B shows a false positive prediction in the non-tumoral parenchyma with a high confidence level (75%). In Figure 4B, the partially obscured lesion was detected with a lower confidence level (42%). Note: Red bounding boxes represent the manually annotated ground truth, and blue bounding boxes are the HCC nodules predicted by the deep learning model.