Preoperative Prediction of Pathological Grading of Hepatocellular Carcinoma Using Machine Learning-Based Ultrasomics: A Multicenter Study

Purpose: The present study investigated the value of ultrasomics signatures in the preoperative prediction of the pathological grading of hepatocellular carcinoma (HCC) via machine learning. Methods: A total of 193 patients were collected from three hospitals. The patients from two hospitals (n = 160) were randomly divided into training set (n = 128) and test set (n = 32) at a 8:2 ratio. The patients from a third hospital were used as an independent validation set (n = 33). The ultrasomics features were extracted from the tumor lesions on the ultrasound images. Support vector machine (SVM) was used to construct three preoperative pathological grading models for HCC on each dataset. The performance of the three models was evaluated by area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy. Results: The ultrasomics signatures extracted from the grayscale ultrasound images could successfully differentiate between high- and low-grade HCC lesions on the training set, test set, and the independent validation set (p < 0.05). On the test set and the validation set, the combined model's performance was the highest, followed by the ultrasomics model and the clinical model successively (p < 0.05). Their AUC (along with 95 %CI) of these models was 0.874(0.709-0.964), 0.789(0.608-0.912), 0.720(0.534-0.863) and 0.849(0.682-0.949), 0.825 (0.654-0.935), 0.770(0.591-0.898), respectively. Conclusion: Machine learning-based ultrasomics signatures could be used for noninvasive preoperative prediction of pathological grading of HCC. The combined model displayed a better predictive performance for pathological grading of HCC and had a stronger generalization ability.