A Non-Parametric Analysis to Estimate Dose Distributions Associated With High-Risk vs. Low-Risk of Death in RTOG 0617.

Researchers

Journal

Modalities

Models

Abstract

Analyses of locally-advanced non-small cell lung cancer (LA-NSCLC) series have identified radiation dose to the cardio-pulmonary system as a new key detriment for risk of death. However, spatial dose patterns are typically represented as one-dimensional dose metrics of pre-defined segmented structures. Here we introduce the use of a recently developed non-parametric approach based on an extended form of optimal mass transport (OMT), called unbalanced OMT distance. Thoracic dose distributions were ranked in relation to risk of death within RTOG 0617. This agnostic approach preserves the spatial dose information without pre-selection of specific thoracic sub-regions or normalization.The 409 patients with evaluable data (CTs, dose, structures and overall survival status) were retrieved from The Cancer Imaging Archive. Patients were censored at 60 months. Data was split into a 70% training and a 30% validation data (n = 286 and 123). Dose distributions were deformably registered to randomly selected reference patients using multi-step B spline. The unbalanced OMT distance measures dose spread pattern and total amount of dose differences among patients. Using a newly proposed approximation of the unbalanced OMT distance via vector-valued OMT, the highly complicated distances between pairs of dose distributions were computed efficiently. The OMT distances were translated into risk scores represented by C-indices: small distances (low C-indices) represent high-risk; large distances (high C-indices) represent low-risk. Risk status in the validation data was measured as distances to the training reference patients. Lastly, two sub-regions were deduced: One high-risk region based on the 50 riskiest patients (overlap of at least 80% of distributions including the highest 80% dose relative to max dose) and one low-risk region based on the 50 least risky patients. The locations of the sub-regions were qualitatively compared to those of eight deep learning auto-segmented cardio-pulmonary substructures (aorta, atria, inferior and superior vena cava (SVC), pulmonary artery (PA), ventricles).The highest risk of death was identified in a region overlapping with the base of the heart, including the ascending aorta, PA and the superior part of the left atrium. In contrast, the ‘safe’ distributions associated with the lowest risk of death mostly avoided all cardio-pulmonary substructures with the exception of the SVC and right PA.This non-parametric approach identified specific high-risk and low-risk regions associated with death in RTOG 0617. While there was no one-to-one correspondence between the identified high-risk region and the segmented cardio-pulmonary structures, this region focused on the base of the heart. The identified regions may direct further work on elucidating mechanisms and generalizability.Copyright © 2021. Published by Elsevier Inc.