Abstract: Radiation-induced liver disease constitutes a major dose-limiting factor for hepatocellular carcinoma (HCC) radiotherapy. As per the recently published ASTRO clinical practice guideline for HCC, the selection of dose-fractionation regimen and modality should consider tumor location and underlying liver function, with a strong recommendation for dose-escalated hypofractionated regimens whenever possible. Here, we developed a model to predict the decline in normal liver function after both photon and proton therapies on a patient-specific basis. The model uses a convolution neural network to extract dosimetric features from the differential dose-volume histogram, interact with baseline measures of liver function and blood biomarkers, and classify the compounded features into post-treatment hepatic toxicity positive and negative classes. We opted for a compact network to maintain a balance of accuracy and interpretability. We trained the model using 117 HCC patients from Massachusetts General Hospital and independently validated it on 88 MD Anderson Cancer Center patients after registering a study analysis plan. Predictive performance and decision-curve analysis indicated a potential utility for individualizing the selection of modality, dose, and fractionation of radiotherapy in HCC.

About the speaker: Ibrahim Chamseddine is a Postdoctoral Fellow in the Department of Radiation Oncology at Massachusetts General Hospital (MGH), Harvard Medical School. His work focuses on cancer outcome prediction and optimization. He obtained his Ph.D. in Mechanical Engineering from McGill University, then attended Moffitt Cancer Center as a Postdoctoral Fellow in Integrated Mathematical Oncology before joining MGH. He is a recipient of the Prince Alwaleed Bin Talal Research Fellowship, GERAD Best Ph.D. Thesis Award, Moffitt Junior Researcher Partnership Award, NCI People’s Choice Award, and Mathematical Oncology Tumor Board Evolution Award.