Development of Prediction Model for Right and Left Liver Hypertrophy to Estimate Response in Patients' Treatments Simulated with Photon Based 3DCRT and IMRT Plans

• Published in: International journal of radiation oncology, biology, physics Vol. 120; no. 2; pp. S217 - S218
• Main Authors: Gupta, A.C., Taie, M. Al, Vazquez, I., Castelo, A., Flint, D.B., Tang, T.T., Prasad, A.S., Zha, Y., Robles, S., O'Connor, C., Leone, A.O., Court, L.E., Yang, M., Yedururi, S., Wu, J., Long, J.P., Howell, R.M., Koay, E.J., Brock, K.K.
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01-10-2024
• ISSN: 0360-3016
• DOI: 10.1016/j.ijrobp.2024.07.2298

Liver hypertrophy after radiotherapy (RT) is poorly understood but may improve survivorship by maximizing the healthy liver tissue post-RT. We developed machine learning (ML) models to predict liver hypertrophy post-RT and estimated/compared liver hypertrophy in 3DCRT versus intensity modulated RT (IMRT) photon plans.
205 patients receiving RT to liver were collected and divided into two cohorts. In all patients, liver segments (1, 2, 3, 4, 5-8), tumor/metastases/cysts were contoured using three AI models on the RT planning and 3-month-followup CT scans. Hypertrophy was binarized by thresholding volumetric increase>7.6cc and 8.5cc for left and right liver. In cohort-1,120 clinical/dosimetric features of patients were filtered using Chi-squared/Fisher-Exact test and Logistic regression. Hypertrophy prediction models for right and left liver were trained using Lasso regression, support vector machine, random forest, XGBoost, Gaussian Process, and Naïve Bayes with train/test split of 148/37 with 10-fold cross validation. Models were assessed via AUROC, sensitivity/specificity, and accuracy. Majority voting (MV) was applied on all models to determine final predictions. In cohort-2, an in-house AI model was used to construct IMRT doses on 20 independent patients who originally received 3DCRT photon treatments. Dosimetric features from both RT groups were separately input to the best hypertrophy prediction models obtained from cohort-1.
Mean Hounsfield Unit (as a surrogate for steatosis) and right liver volume spared from 25GyEQD2 were significant(p<0.05) predictors of right liver hypertrophy. Tumor location, minimum dose to 95% of GTV, mean doses to right and left liver volume spared from 40GyEQD2 were significant predictors of left liver hypertrophy(p<0.05). In cohort-1, XGBoost showed the best scores with AUROC, sensitivity, specificity, and accuracy of 0.72, 0.79, 0.65, 0.68 for right liver. For left liver, MV showed best AUROC, sensitivity, specificity, and accuracy of 0.65, 0.79, 0.50, 0.65, respectively. In cohort-2, for right liver hypertrophy, 3/20 patients showed hypertrophy with both RT plans, 2/20 showed hypertrophy for only 3D, and 2/20 showed hypertrophy for only IMRT. For left liver hypertrophy, 5/20 patients showed hypertrophy using IMRT only, but no patients showed hypertrophy using 3D-plan only. 13/20 patients showed hypertrophy using both plans.
Tumor location and liver volume spared from specific RT doses are significant predictors of hypertrophy. ML can describe the liver hypertrophy with reasonable accuracy. Liver hypertrophy varies with dose distribution, which signals RT plans can be optimized to maximize liver hypertrophy.