AI-Driven Segmentation and Volumetric Response Modeling of Liver Regions to Radiotherapy

Aashish Chandra Gupta, MS (Advisor: Kristy K. Brock, PhD)

In liver-directed radiotherapy (RT), liver regions receiving higher doses typically undergo atrophy while contralateral/adjacent lower-dose regions may exhibit compensatory hypertrophy through regeneration of healthy tissue. Optimizing the RT plan to promote regional hypertrophy while minimizing the risk of developing atrophy has the potential to enhance post-RT liver function and long-term survivorship. However, current clinical practice largely relies on global liver dose-volume metrics during RT-planning, which may obscure favorable dose-response correlation and limit actionable guidance for clinicians. Therefore, we hypothesized that post-RT regional liver response is governed by a combination of region-specific dose-volume and patient clinical features, and that these responses can be predicted with clinically meaningful accuracy using AI-driven models to guide RT technique selection.

To address this, we developed fully automated deep-learning-based auto-contouring models using nnU-Net and patch-based Attention U-Net (paU-Net) utilizing 200 CT images to auto-contour liver segments 1, 2, 3, 4, 5-8 and spleen. nnU-Net architecture outperformed paU-Net, achieving a final mean Dice of 0.89, 0.82, 0.88, 0.87, 0.96, and 0.95 for segments 1, 2, 3, 4, 5–8, and spleen, respectively on test sets. Qualitative analysis showed over 85% of test cases achieved a Likert score>3, indicating clinically acceptable contours.

Using the auto-contouring model, we curated a cohort of 207 patients treated with RT for diverse liver tumor types and quantified an uncertainty threshold of ±15.8cc to define clinically meaningful functional regional-volume change (ΔFV)  at three-months post-RT across 561 liver regions. Hypertrophy (ΔFV > 15.8cc), atrophy (ΔFV < - 15.8cc), and stable responses were observed in 28%, 50%, and 22% of regions respectively. A mean dose threshold of 15Gy_EQD2 distinguished hypertrophy from non-hypertrophy response with an area under receiver operating characteristic curve (AUROC) of 0.71; however, the threshold did not generalize when assessing rate of hypertrophy/atrophy below 15Gy_EQD2 across tumor types and liver regions which supported the need for region-specific predictive modelling.

We subsequently developed and validated machine learning models predicting hypertrophy and atrophy responses of right, left lateral, and left central liver regions. Cross-validation AUROCs ranged from 0.66 to 0.76. Key features included albumin-bilirubin score, BMI, functional volume, percent volume of a region spared from 5-50Gy_EQD2 normalized to body-surface area, and minimum doses received by more than 2%, 50% and 98% of liver regions. On a withheld test set, accuracy for right, left lateral, left central hypertrophy/not-hypertrophy models were 0.77, 0.72, 0.1 and for atrophy/not-atrophy were 0.63, 0.68, 0.72, respectively. Model interpretability analysis revealed a strong association between regional dose sparing and hypertrophy, whereas higher point dose such as D2 were associated with atrophy.

Finally, we demonstrated the potential clinical utility of our models by comparing 3DCRT and VMAT plans. Although both treatment approaches exhibited low overall rate of regional-hypertrophy, VMAT resulted in hypertrophy in 4 out of 12 liver regions compared with 2 out of 12 regions for 3DCRT, signaling patient-specific advantage of RT technique.

In conclusion, our work quantified clinically meaningful relations between radiation dose and regional liver response post-RT, and demonstrated the utility of AI-driven segmentation and regional response modelling for personalization of liver-directed RT.

Advisory Committee:

• Kristy K. Brock, PhD, Chair
• Rebecca M. Howell, PhD
• Eugene J. Koay, MD, PhD
• James P. Long, PhD
• Jian Wu, PhD

Join via Zoom (Please contact Mr. Gupta for his Zoom meeting info.)