ESTRO 2024 - Abstract Book

S4571

Physics - Machine learning models and clinical applications

ESTRO 2024

3143

Digital Poster

Deep learning improves registration accuracy for voxel-based radiotherapy outcomes modeling

Chloe Min Seo Choi 1,2 , Jue Jiang 1 , Joseph O. Deasy 1 , Andreas Rimner 3 , Maria Thor 1 , Harini Veeraraghavan 1

1 Memorial Sloan Kettering Cancer Center, Department of Medical Physics, New York, USA. 2 Yonsei University College of Medicine, Department of Radiation Oncology, Seoul, Korea, Republic of. 3 Memorial Sloan Kettering Cancer Center, Department of Radiation Oncology, New York, USA

Purpose/Objective:

Inter-patient image registration is a process of normalizing multiple patients into the same reference plane [1]. This registration process ensures that images are aligned at the voxel level, thereby enabling applications that require voxel-based comparisons including voxel-based analysis (VBA) for outcomes modeling. Most VBA approaches use iterative deformable image registration (DIR) methods to align patients. These methods mask out anatomically different image portions and pre-select patients without large tumours for analysis to avoid excluding many patients from poor registration accuracy. This requires manual preprocessing. Hence, we evaluated whether a deep learning (DL)-based approach [2] was more accurate than an iterative DIR method for enabling automated VBA while considering gender-specific differences in registration accuracy.

Material/Methods:

A dataset with 22 male and 18 female patients with non-small cell lung cancer was analyzed. We selected one representative male and female patient based on the median body mask size. The remaining patients were aligned to each reference using DL-based registration in parallel to a conventional iterative image registration method called symmetric diffeomorphic image registration (SyN). Registration accuracy was judged using the Dice Similarity Coefficient (DSC) for the lung in comparison to the reference patient. Subsequently, patients whose lung DSC was < 0.80 were excluded. We stratified the analysis according to the gender to study its influence on registration accuracy. Statistical comparisons were performed using Wilcoxon signed-rank test.

Results:

Our DL-based method achieved a significantly higher DSC than SyN (Table 1). Additionally, our method had one patient exclusion due to DSC < 0.80, while SyN had 19 and 23 exclusions using male and female references, respectively. Male vs. female results showed slightly lower performance when the male reference was used. However, stratified analysis showed that DSC produced with the DL method was not dependent on gender. SyN showed significantly higher DSC in female patients when using the female reference, but a study with a larger patient cohort is necessary to confirm this result.

No.

of

excluded

p-value

(male

vs

DSC

patients

female DSC)