Differentiation of focal liver lesions in contrast-enhanced ultrasound using a heuristic-guided hybrid machine-learning framework.

[PURPOSE] Sonazoid contrast-enhanced ultrasound (CEUS) offers valuable diagnostic information on hepatic lesions, but it is time-consuming. In this study, we investigated a novel composite machine-learning framework that integrates heuristic knowledge and model-specific classification to differentiate liver lesions using only the first 2 min of CEUS imaging.

[METHODS] CEUS images from 232 patients with 232 focal liver lesions (benign: 61, hepatocellular carcinoma [HCC]: 104, non-HCC malignancies [ML]: 67) were analyzed. For each case, six frames from injection to peak enhancement and static images at 1 and 2 min post-injection were used. Two deep learning models were developed: Model 1 classified heterogeneous enhancement patterns into "benign," "HCC," "ML," or "Uniform" (homogeneous). Model 2 further classified "Uniform" cases into three diagnostic categories. Lesion brightness values were incorporated as input features. The artificial intelligence (AI) mode was also evaluated by observer study of three hepatologists using the area under the receiver operating characteristic curve (AUC).

[RESULTS] The composite model was evaluated on 58 independent test cases, achieving classification accuracy of 81.8% for benign, 93.5% for HCC, and 68.8% for ML, with an overall accuracy of 84.5%. Binary classification (benign vs. malignant) yielded 97.9% sensitivity, 94.8% specificity, and 94.8% overall accuracy. For discrimination between benign and malignant, the mean AUC for the three observers was significantly improved with AI output, where the difference in AUC (95% confidence interval) was 0.095 (0.0197, 0.1703) (P = 0.013).

[CONCLUSIONS] The proposed AI-based framework enables accurate liver lesion classification using early phase CEUS, eliminating the need for Kupffer-phase imaging in many cases.