Deriving Pulmonary Perfusion Images from 4DCT Using Deep Learning

Document Type

Conference Proceeding

Publication Date

3-2021

Publication Title

Journal of Thoracic Oncology

Abstract

Introduction

The current standard for ventilation and perfusion imaging, SPECT/CT, is accurate but costly and resource intensive, thus alternative approaches warrant investigation. A deep learning-based perfusion imaging approach, which forgoes the need for SPECT imaging and instead uses 4DCT, could expand the accessibility of functional avoidance radiation therapy for lung cancer patients. Previous literature by Guerrero et al. has shown that lung mass variation across the phases of a 4DCT imaging study can be used to derive pulmonary perfusion during tidal breathing. In this study, we investigate the derivation of SPECT perfusion imaging from only the maximum inhalation and exhalation phases of 4DCT images.

Methods

This retrospective study utilized 26 lung cancer patients who underwent both 4DCT and SPECT/CT imaging sessions at two time points on a clinical trial (R01CA200817). The images from these two modalities were registered using MIM software. Both the CT and SPECT perfusion images were then cropped to the lung regions alone. The cropped 4DCT images were used as input and the cropped SPECT perfusion images as the ground truth. The dataset of 26 patients with 51 imaging studies was split by patient into 41 / 5 / 5 studies for the training, validation and testing sets. Using the training and validation sets, a High-Res3DNet was trained and hyperparameter tuned. Final results were reported from the hold-out testing data set.

Results

Gamma analysis of 5% / 5mm and Spearman's Rank Correlation Coefficient were used to compare SPECT perfusion images and deep learning predictions. For visual comparison, a figure is provided.

Conclusion

Our results perform similarly to a previous investigation which utilized derived perfusion images from 4DCT, as opposed to the clinically acquired SPECT perfusion studies. With an improved model design, we expect a consistent diagnostic alternative to SPECT images.

Volume

16

Issue

3 (Supplement S)

First Page

S137

DOI

10.1016/j.jtho.2021.01.213

Link to Full Text

Share

COinS