CT-Ventilation Using Swin UNETR Networks for Automated Lobe Segmentation

Document Type

Conference Proceeding

Publication Date

10-2024

Publication Title

International Journal of Radiation Oncology, Biology, Physics

Abstract

Purpose/Objective(s): CT-ventilation methods typically rely on deformable image registration (DIR) to calculate the apparent volume changes within an inhale/exhale computed tomography (CT) image pair. However, DIR is itself a non-trivial task that can potentially bias the resulting CTventilation. We hypothesize that the breathing-induced lung volume changes needed to calculate CT ventilation can be accurately computed directly from AI driven lobe segmentations, without a DIR. We utilize a novel transformer-based lobe segmentation pipeline that is comprised of three Swin UNETR networks. The resulting inhale/exhale lobe volumes are then used to calculate CT-ventilation. Materials/Methods: Our segmentation model is based on the Swin UNETR network. Our segmentation pipeline involves the application of three Swin UNETR networks. Each network was trained on 6,501 unlabeled inhale/exhale CT image pairs from the COPDgene imaging study. First, an initial network provides right and left lung segmentations that are used to define bounding boxes for each lung. The bounding boxes are resized to focus on the lung volumes and then lobes are segmented with dedicated right and left lung Swin UNETR networks. Fine-tuning was conducted on manual lobe segmentations from diverse clinical datasets; 40 lung nodule cases from the LUNA16 challenge and CT from 11 patients treated with radiotherapy for non-small cell lung cancer. Segmentation accuracy was assessed on 10 LUNA16 test cases with manually delineated ground truth. Breathing-induced volume change (exhale volume/inhale volume) was calculated for each lobe using the Swin UNETR-defined lobe volumes delineated on the inhale/exhale phases, without the use of DIR. The resulting lobar CT-ventilation values were validated using publicly available 4DCT and PET-Galligas ventilation imaging for 19 lung cancer patients. The spatial Spearman correlation between our Swin UNETR lobe ventilation and the ground-truth PET Galligas images was calculated for each patient. Results: The segmentation pipeline achieved a mean Dice score of 0.9538 (LUL = 0.9725, LLL = 0.9700, RUL = 0.9574, RML = 0.9063, and RLL = 0.9626) on the LUNA16 test cases, surpassing the mean Dice accuracies reported by current state-of-the-art models. CT-ventilation calculations yielded a median Spearman correlation coefficient of 0.8 across 19 cases, with 14 cases exhibiting correlations of at least 0.5, indicating strong agreement with PET Galligas ventilation. Upon visual inspection, lower correlations were associated with low 4DCT image quality. Conclusion: Our Swin UNETR lobe segmentation pipeline demonstrated superior performance over prior methods, while our segmentation-based CT ventilation method exhibits strong agreement with PET Galligas ventilation. Moreover, as our approach leverages deep learning for segmentation, it provides interpretable ventilation results and facilitates straightforward quality assurance processes, thereby reducing reliance on DIR-dependent algorithms.

Volume

120

Issue

2S

First Page

S52

Comments

ASTRO 2024: 66th Annual Meeting American Society for Radiation Oncology, September 29 - October 2, 2024, Washington, DC

DOI

10.1016/j.ijrobp.2024.07.085

Share

COinS