Predictive Model of Liver Toxicity to Aid the Personalized Selection of Proton Versus Photon Therapy in Hepatocellular Carcinoma

Abstract

Purpose: Our objective was to develop an externally validated model for predicting liver toxicity after radiation therapy in patients with hepatocellular carcinoma (HCC) that can integrate both photon and proton dose distributions with patient-specific characteristics.

Methods and materials: Training data consisted of all patients with HCC treated between 2008 and 2019 at our institution (n = 117, 60%/40% photon/proton). We developed a shallow convolutional neural network (CNN) to predict posttreatment liver dysfunction from the differential dose-volume histogram (DVH) and baseline liver metrics. To reduce bias and improve robustness, we used ensemble learning (CNNE). After a preregistered study analysis plan, we evaluated stability using internal bootstrap resampling and generalizability using a data set from a different institution (n = 88). Finally, we implemented a class activation map method to characterize the critical DVH subregions and benchmarked the model against logistic regression and XGBoost. The models were evaluated using the area under the receiver operating characteristic curve and area under the precision-recall curve.

Results: The CNNE model showed similar internal performance and robustness compared with the benchmarks. CNNE exceeded the benchmark models in external validation, with an area under the receiver operating characteristic curve of 0.78 versus 0.55 to 0.70, and an area under the precision-recall curve of 0.6 versus 0.43 to 0.52. The model showed improved predictive power in the photon group, excellent specificity in both modalities, and high sensitivity in the photon high-risk group. Models built solely on DVHs confirm outperformance of the CNNE and indicate that the proposed structure efficiently abstracts features from both proton and photon dose distributions. The activation map method demonstrates the importance of the low-dose bath and its interaction with low liver function at baseline.

Conclusions: We developed and externally validated a patient-specific prediction model for hepatic toxicity based on the entire DVH and clinical factors that can integrate both photon and proton therapy cohorts. This model complements the new American Society for Radiation Oncology clinical practice guidelines and could support value-driven integration of proton therapy into the management of HCC.

Fig. 1.

CNN modeling. (A) Example dDVHs from the MGH data set showing the differences between photon and proton therapy dose distributions. (B) Structure of the shallow CNN model used for predicting CP2+, composed of feature extraction and feature interaction stages. A dropout rate of 0.5 is used. (C) Ensemble modeling approach used to stabilize the model, where 5 submodels (CNN i) of the same structure are trained on balanced bootstrapped subsets of the data, and their prediction (P_i) is averaged. Abbreviations: ALB = albumin; BIL = bilirubin; CNN = convolutional neural network; CP = Child-Pugh; dDVH = differential dose-volume histogram; EQD2 = equivalent dose in 2 Gy fractions; FC = fully connected layer; GTV = gross tumor volume; MLD = mean liver dose; PLT = platelet.

Fig. 2.

Model evaluation. Internal and external validation curves of the (A) benchmark logistic regression model and (B) main CNNE model. The first column lists the variables used by each model, the second column shows the internal ROC generated using 1000 bootstrap iterations of the MGH data set, mean AUC and 95% confidence interval. The third and fourth columns displays the external ROC and PRC of each model, tested on MDACC data set, showing the area under each curve, AUC and AUPRC, respectively. Abbreviations: AUC = area under the curve; AUPRC = area under the precision-recall curve; CI = confidence interval; CNN = convolutional neural network; CNNE = convolutional neural network ensemble; EQD2 = equivalent dose in 2 Gy fractions; FC = fully connected layer; PRC = precision-recall curve; ROC = received operating characteristic.

Fig. 3.

Comparison of models’ external performance. (A) AUC and (B) AUPRC of the main CNNE and the benchmark models: CNN, LR, XGB, as well as all the algorithms using dose-only features, that is, without liver biomarkers. The ROC and PRC of each model can be found in Figs. E3–6 and E8. The x-axis of the ROC is shifted to 0.5, the reference to the random guess, and for the PRC to 0.35, the imbalance of data in MDACC. Abbreviations: AUC = area under the curve; AUPRC = area under the precision-recall curve; CNN = convolutional neural network; CNNE = convolutional neural network ensemble; LR = logistic regression; PRC = precision-recall curve; ROC = received operating characteristic; XGB = XGBoost.

Fig. 4.

Result of activation map method showing the risk incurred by different dose-volume histogram regions in patients with baseline low liver function (CP = 6+) relative to normal liver function (CP = 5), showing an increasing criticality of the low dose region as the baseline CP increases. Abbreviation: CP = Child-Pugh; EQD2 = equivalent dose in 2 Gy fractions.

Fig. 5.

Model-driven decision support map for personalized selection of treatment modality based on the predicted liver toxicity.