Radiomics-based prediction of HCC response to atezolizumab/bevacizumab

Abstract

Advanced hepatocellular carcinoma (HCC) treatment has evolved with the introduction of atezolizumab/bevacizumab, showing improved outcomes over sorafenib. However, the response varies among patients, particularly between viral and non-viral etiologies. The present study aimed to develop and evaluate multimodal prediction models combining quantitative imaging and clinical markers to predict the treatment response in patients with HCC. Between March 2020 and May 2023, patients with advanced HCC treated with atezolizumab/bevacizumab were retrospectively identified from six centers in Germany and Austria. Patients underwent baseline contrast-enhanced liver MRI and follow-up imaging to assess the therapy response. Machine learning models, including RandomForestClassifier, were developed for radiomics, clinical and combined datasets. Hyperparameter tuning was performed using RandomizedSearchCV, followed by cross-validation to evaluate model performance. The study included 103 patients, with 70 achieving disease control (DC) and 33 experiencing disease progression (PD). Key findings included significant differences in treatment response and progression-free survival between the DC and PD groups. The radiomics model, using 14 selected features, achieved 73.1% accuracy and a receiver operating characteristic (ROC) area under the curve (AUC) of 0.635 for the test set. The clinical model, with 4 selected features, achieved 73% accuracy and a ROC AUC of 0.649 for the test set. The combined model showed improved performance, with 69% accuracy and a ROC AUC of 0.753 for the test set. Hyperparameter tuning further enhanced the accuracy of the combined model to 80.1% and the ROC AUC to 0.771 for the test set. In conclusion, the hybrid model combining clinical and radiological data outperformed individual models, providing improved predictions of response to atezolizumab/bevacizumab in patients with HCC.

Keywords: checkpoint inhibitors; hepatocellular carcinoma; machine learning; prediction models; radiomics.

Figure 1.

Overall and progression-free survival. Survival curves: (A and C) Overall and (B and D) progression-free survival according to (A and B) control vs. progression and (C and D) modified Response Evaluation Criteria on Solid Tumors category. CR, complete response; PD, disease progression; PR, partial response; SD, stable disease.

Figure 2.

Radiomics features represented as a heatmap in the ROI. Example of a radiomics heatmap in (A) a 62-year-old male patient who responded to therapy and survived for at least 484 days after the beginning of atezolizumab and bevacizumab treatment and (B) a 70-year-old male patient who progressed in spite of therapy but still survived at least 266 days after the beginning of immunotherapy. The ROI corresponds to a hepatocellular carcinoma tumor and the radiomics features depicted are: Large area low gray level emphasis, which highlights regions of large, low-intensity areas, with high values indicating homogeneous, low-contrast zones; and zone variance, which measures the variability in zone sizes, where high values suggest greater heterogeneity in texture, while low values reflect more uniform zone sizes. ROI, region of interest.

Figure 3.

Feature importance in the different models. (A) Radiomics model. 1, GLSZM large area low gray level emphasis; 2, GLCM cluster shade; 3, GLSZM zone variance; 4, GLSZM large area high gray level emphasis; 5, GLCM joint average; 6, first order skewness; 7, GLSZM large area emphasis; 8, GLRLM short run high gray level emphasis; 9, GLRLM long run low gray level emphasis; 10, GLCM autocorrelation; 11, GLCM sum average; 12, GLSZM low gray level zone emphasis; 13, GLRLM high gray level run emphasis; and 14, GLDM high gray level emphasis. (B) Clinical model. 1, C-reactive protein; 2, metastasis status; 3, previous systemic therapy; and 4, previous TACE. (C) Combined model. 1, GLSZM zone variance; 2, GLSZM large area high gray level emphasis; 3, GLRLM high gray level run emphasis; 4, GLDM high gray level emphasis; 5, GLRLM long run low gray level emphasis; 6, GLCM joint average; 7, GLCM autocorrelation; 8, first order skewness; 9, GLRLM short run high gray level emphasis; 10, GLSZM low gray level zone emphasis; 11, previous systemic therapy; 12, metastasis status; 13, GLCM sum average; 14, C-reactive protein; 15, GLSZM large area emphasis; 16, GLSZM large area low gray level emphasis; 17, GLCM cluster shade; and 18, previous TACE. GLCM, gray level co-occurrence matrix; GLDM, gray level dependence matrix; GLRLM, gray level run length matrix; GLSZM, gray level size zone matrix; TACE, transarterial chemoembolization.

Figure 4.

ROC curves for each of the different models. ROC curves for the (A) radiomics, (B) clinical and (C) combined models. ROC, receiver operating characteristic.

Figure 5.

ROC curves after hyperparameter tuning. ROC curves of the (A) training set and (B) testing set after hyperparameter tuning. For the combined model, the AUC was 0.77 in the testing set, with an accuracy of 80.1%, sensitivity of 62.5% and specificity of 89%. AUC, area under the curve; ROC, receiver operating characteristic.