Machine learning model based on dynamic contrast-enhanced ultrasound assisting LI-RADS diagnosis of HCC: A multicenter diagnostic study

Abstract

Background: To enhance the accuracy of hepatocellular carcinoma (HCC) diagnosis using contrast-enhanced (CE) US, the American College of Radiology developed the CEUS Liver Imaging Reporting and Data System (LI-RADS). However, the system still exhibits limitations in distinguishing between HCC and non-HCC lesions.

Purpose: To investigate the viability of employing machine learning methods based on quantitative parameters of contrast-enhanced ultrasound for distinguishing HCC within LR-M nodules.

Materials and methods: This retrospective analysis was conducted on pre-treatment CEUS data from liver nodule patients across multiple centers between January 2013 and June 2022. Quantitative analysis was performed using CEUS images, and the machine learning diagnostic models based on quantitative parameters were utilized for the classification diagnosis of LR-M nodules. The performance of the model was assessed using the area under the receiver operating characteristic curve (AUC) and compared with the performance of four radiologists.

Results: The training and internal testing datasets comprised 168 patients (median age, 53 years [IQR, 18 years]), while the external testing datasets from two other centers included 110 patients (median age, 54 years [IQR, 16 years]). In the internal independent test set, the top-performing Random Forest model achieved an AUC of 0.796 (95%CI: 0.729-0.853) for diagnosing HCC. This model exhibited a sensitivity of 0.752 (95%CI: 0.750-0.755) and a specificity of 0.761 (95%CI: 0.758-0.764), outperforming junior radiologists who achieved an AUC of 0.619 (95%CI: 0.543-0.691, p < .01) with sensitivity and specificity of 0.716 (95%CI: 0.713-0.718) and 0.522 (95%CI: 0.519-0.526), respectively.

Conclusion: Significant differences in contrast-enhanced ultrasound quantitative parameters are observed between HCC and non-HCC lesions. Machine learning models leveraging these parameters effectively distinguish HCC categorized as LR-M, offering a valuable adjunct for the accurate classification of liver nodules within the CEUS LI-RADS framework.

Keywords: Dynamic contrast-enhanced ultrasound; Hepatocellular carcinoma; Image enhancement; Liver imaging reporting and data system; Machine learning; Quantitative ultrasound; Time-intensity curve; Ultrasonography.

Fig. 1

Flow diagram of the study. ∗The classification criteria were based on CEUS Liver Imaging Reporting and Data System (LI-RADS) version 2017. HCC = hepatocellular carcinoma, non-HCC = Nodules excluding HCC.

Fig. 2

ROI selection and TIC of different lesions. (A, B) Images of a 68-year-old man displaying a nodule classified as LR-M and confirmed as hepatocellular carcinoma through surgery. (A) A 4.2 × 3.2 cm hypo-echoic nodule (arrowheads) in segment 6 of the right liver lobe is evident on the surveillance ultrasound (US) image, demonstrating early wash-out (<60 s; timer, 00:59) but presenting mild washout in the contrast-enhanced (CE) US images of the nodule (highlighted by the green box). (B) Representative curves of HCC derived from quantitative analysis software. (C, D) Images of a 62-year-old woman showing a nodule classified as LR-M and confirmed as liver metastases biopsy. (C) A 7.2 × 4.6 cm hypo-echoic nodule (arrowheads) in segment 7 of the right liver lobe is observed on the surveillance US image, indicating early wash-out (<60 s; timer, 00:48) but more marked washout in the CEUS images of the nodule (highlighted by the green box). (D) Typical curves of non-HCC lesions obtained from quantitative analysis software. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

Fig. 3

The definitions of the quantitative parameters (The data in this example has been normalized). The X axis is the time, and the Y axis is the enhancement intensity. The TIC of the target tissues is shown in green, and the liver parenchyma is shown in yellow, where the elongated curve corresponds to the original video signal intensity and the re-curve is fitted. The linearized signal displays echo-power data over time, showing how the intensity of the reflected echoes changes as a function of time. AT: Arrival time; PI: Peak intensity; PT: Peak time; PX: Time at the intersection of the TIC; WiMTT: Mean transit time of the wash-in segmen; WoMTT: Mean transit time of the was-out segment. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)

Fig. 4

Diagnostic performance comparison between the ML models and radiologists in HCC identification with liver nodules. (A) Diagnostic performance of radiologists and ML models compared within the internal independent testing set. (B) Diagnostic performance of radiologists and ML models compared within the external independent testing set. Circles represent the diagnostic sensitivity and specificity of the senior radiologist, while triangles indicate the sensitivity and specificity of the junior radiologist. ML = machine learning. AUC = area under the receiver operating characteristic curve.

Fig. 5

SHapley Additive exPlanations (SHAP) Summary and Dependency Plots for Key Features. (A) The SHAP summary plot. Each dot represents the SHAP value of a specific feature for an individual patient, with multiple dots converging into a violin plot where the width indicates patient density. The color gradient, ranging from blue to red, reflects the variation in feature values from low to high. The Y-axis lists the features, ordered by their average impact on the model's predictions, while the X-axis represents the SHAP value's influence on the model's output. A larger SHAP value correlates with a higher probability of HCC, with positive values indicating a positive impact on the model's prediction and negative values indicating a negative impact. (B–D) The SHAP dependency plots. (B) PX_rate vs. WoMTT_A. (C) WiAUC_A vs. WoMTT_A. (D) WoMTT_A vs. PX_rate. The X-axis represents the feature values, while the Y-axis shows the corresponding SHAP values, indicating each feature's contribution to the model output. Each point's color, ranging from red to blue, represents the interacting feature's value from high to low. This color gradient typically reflects the relationship between the primary and interacting features, as well as their combined influence on the model's prediction. (For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)