3014 - Machine Learning-Based Prediction of Regional Liver Volume Response after Radiotherapy in Patients with Liver Malignancies
Presenter(s)
A. C. Gupta1, M. Altaie2, T. T. Tang1, A. Castelo1, C. O'Connor1, D. B. Flint3, S. Yedururi4, M. B. Saad1, E. J. Koay5, and K. K. Brock3; 1Department of Imaging Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, 2UT MD Anderson Cancer Center, Houston, TX, 3Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, TX, 4Department of Abdominal Imaging, The University of Texas MD Anderson Cancer Center, Houston, TX, 5Department of Radiation Oncology, The University of Texas MD Anderson Cancer Center, Houston, TX
Purpose/Objective(s): Understanding how different liver regions respond to radiotherapy (RT), while considering clinical factors that affect baseline liver function, may help clinicians optimize RT planning. We aimed to develop region-specific models to predict liver hypertrophy and atrophy after RT in patients with liver malignancies.
Materials/Methods: A total of 170 patients treated with liver-directed RT were included: hepatocellular carcinoma (HCC, n=47), cholangiocarcinoma (CC, n=71), combined HCC/CC (n=3), and liver metastases (Mets, n=49). Planning and three-month follow-up CT scans were imported. Liver regions—central (segments 1+4), left (segments 2+3), and right (segments 5–8)—were generated using AI and were corrected by radiologists. Statistical and risk analyses based on mean dose were performed prior to model development to evaluate hypertrophy/atrophy by histology and region. Candidate predictors for model training included DVH metrics, induction chemotherapy, underlying liver disease surrogates, and blood biomarkers. The cohort was split into training (n=133) and independent test (n=37) sets, preserving the proportion of hypertrophy/not-hypertrophy and atrophy/not-atrophy cases. Right and left liver models were trained using two-staged cross-validation (outer=5, inner=3) on the training set. Stage 1 involved a selection of top 10 features using Random Forest; stage 2 involved training nine algorithms using selected features. The final model selected during cross-validation was tested on the withheld test set. Performance was assessed using AUROC, accuracy, sensitivity, and specificity.
Results: Overall, a mean dose of 15 Gy classified hypertrophy versus non-hypertrophy across liver regions, with a pronounced increase in rate of hypertrophy below 15 Gy across all regions in the entire cohort. Feature selection showed that response is driven by BMI/BSA, baseline liver function (ALBI, bilirubin, AST/ALT), regional volumes, tumor burden within central liver, and DVH metrics (D2, D50, D98, and percent of liver region volume spared from radiation). For the right liver, the best hypertrophy and atrophy models achieved AUROC of 0.73±0.08 and 0.74±0.11 on cross-validation set, with accuracy, sensitivity, specificity of 0.77, 0.78, 0.77 and 0.63, 0.71, 0.50 on the test set, respectively. For the left liver, hypertrophy and atrophy models achieved AUROC of 0.74±0.06 and 0.76±0.15 on cross-validation set, and accuracy, sensitivity, specificity of 0.72, 0.45, 0.93 and 0.68, 0.50, 0.80 respectively on the test set.
Conclusion: Liver response post-RT is impacted by a complex interplay of dose, chemotherapy, and baseline liver function, with lower radiation dose resulting in increased rate of hypertrophy. Region-specific machine-learning models can predict liver hypertrophy/atrophy after RT, which may support personalized selection of RT approaches.