- This paper describes a patient-specific prediction model for hepatic toxicity in which an ensemble convolutional neural network (CNNE) was trained on dose-volume histograms and clinical data from 117 patients with hepatocellular carcinoma
- In an external validation cohort of 88 patients the CNNE consistently outperformed benchmark models, showing it can efficiently abstract dosimetric features from patients who receive either proton or photon therapy
- The CNNE also demonstrated substantial power in the external cohort to discriminate whether patients did or did not experience hepatic toxicity, both overall (Cohen's d=1.00) and in the photon therapy subgroup (Cohen's d=1.20)
- The research team was able to identify the relative increase in risk of hepatic toxicity for a given dose level according to a patient's pretreatment liver function score
- The team proposes a decision map for personalized treatment selection based on the predicted risk of liver toxicity
Liver-directed radiation therapy (RT) is now a viable treatment option for unresectable hepatocellular carcinoma and as a bridge to liver transplantation. A substantial number of patients present with poor liver reserve, however.
Subscribe to the latest updates from Cancer Center Advances in Motion
Theodore S. Hong, MD, director of Gastrointestinal Radiation Oncology and co-director of the Tucker Gosnell Center for Gastrointestinal Cancers at the Mass General Cancer Center, Ibrahim Chamseddine, PhD, research fellow, and colleagues are pioneering the use of artificial intelligence to predict hepatic toxicity after either photon or proton RT.
Their latest method, adapted from previous work in lung cancer by a University of Michigan team, relies on using convolutional neural networks (CNNs). This form of artificial intelligence is trained on a type of image and learns how to interpret additional images, in this case dose-volume histograms (DVHs).
The researchers report the development and validation of their model in the International Journal of Radiation Oncology, Biology, Physics.
The researchers trained an "ensemble" of five CNNs (CNNE) on DVHs and clinical data from 117 HCC patients treated with ablative RT between 2008 and 2019. As benchmarks, they developed logistic regression and XGBoost models on standard dosimetric and non-dosimetric patient features.
The task of all models was to predict whether or not patients would experience hepatic toxicity, defined here as a ≥2-point increase in Child–Pugh score from baseline within three months after RT.
On internal validation the AUC was:
- CNNE—0.76 (95% CI, here a measure of stability, 0.61–0.90)
- Logistic regression—0.72 (0.57–0.86)
- XGBoost—0.78 (0.64–0.91)
External validation was performed on data from 88 HCC patients treated at a separate institution between 2006 and 2016:
- CNNE—AUC, 0.78
- Logistic regression—AUC, 0.55
- XGBoost—AUC, 0.68
The CNNE similarly outperformed the benchmark models regarding accuracy in the 10% high- and low-risk groups.
In the external cohort CNNE demonstrated substantial power to discriminate whether patients did or did not experience hepatic toxicity, both overall (Cohen's d=1.00) and in the photon therapy subgroup (Cohen's d=1.20). The odds ratio was 4.2 when splitting patients by the median probability and 6.5 when splitting the top and bottom risk quartiles.
By generating a dose-dependent risk map of the CNNE the research team was able to determine the risk of different dose levels for patients with declining liver function before treatment. Every dose region contributed to the risk of hepatic toxicity, but the low-dose region was more critical than the high-dose region when pretreatment CP score increased above 6.
The paper includes a decision tree for using CNNE results to select the treatment modality (photon vs. proton) and choose whether to decrease the dose of proton therapy.
Learn more about radiation oncology at the Mass General Cancer Center
Learn more about the Mass General Cancer Center