StaticLungNet: A deep learning method to estimate strain from static computed tomography images

Introduction:

Introduction: Lung injuries, such as inflammation ventilator-induced lung injury (VILI), and radiation-induced lung injury (RILI), can lead to heterogeneous alterations in the biomechanical behavior of the lungs [1]. Imaging methods, such as X-ray imaging and static computed tomography (CT) scans, are available for the clinical assessment of lung injuries. While such methods provide morphological information about the lung structure, they cannot directly provide information about the kinematics of lung parenchyma. Image registration has gained recent interest as a post-processing tool to quantify the kinematic behavior of lung parenchyma. Lung image registration is used to estimate the displacement and deformation experienced by the lungs during respiration. However, a limiting factor in the clinical translation of image registration-based mechanistic markers is the need for dynamic computed tomography (CT) images with a sufficiently high temporal resolution corresponding to small displacements between successive timepoints. Dynamic CT scans are avoided in clinical settings due to excessively high radiation. However, clinics often utilize static images acquired at the end-expiratory (EE) and end-inspiratory (EI) timepoints for diagnostic purposes. Thus, a computationally inexpensive method to estimate regional mechanistic biomarkers would greatly enhance the feasibility of their usage in clinics for diagnostic purposes. Thus, in this study, we propose to use a machine-learning (ML) assisted method, trained on synthetically generated dynamic CT images, to predict voxel-level displacement from static CT images acquired at the EE and EI timepoints.

Materials and

Methods: Materials and

Methods: Lung geometries were segmented and reconstructed (Fig. 1A) from the empire10 repository [2]. Then the geometries were meshed using quadratic tetrahedral elements for in-silico simulations in ABAQUS. The material behavior of the lungs was modeled as a neo-Hookean compressible material. Uniform outward forces were applied to the outer surface of the lungs and a uniform downward force was applied to the diaphragm to simulate respiration (Fig. 1B). The mesh at 10 uniformly spaced timepoints in the inspiratory cycle was used to generate in-silico phantoms (Fig. 1C, D)based on the method detailed in our previous work to surrogate dynamic CT image stacks [3]. Image registration was performed on these image stacks to obtain voxel displacement. Then we developed a UNet based ML model to estimate voxel displacement. The input to the model consisted of the first and last image stacks along with their respective segmentations. The voxel displacements obtained from image registration were used as the output of the model.

Results, Conclusions, and Discussions: Results and

Discussion: The in-silico simulation was performed to simulate respiration and 10 in-silico phantoms were generated at equally spaced timepoints from EE to EI (Fig. 1C, D). All 10 phantoms will be used to perform image registration to estimate voxel displacement. These strains will be compared with the strains from the in-silico simulations to ensure accuracy. Then, the first and last image stacks of the in-silico phantoms, their segmentations and the voxel displacements from image registration will be used to train the ML model to predict voxel displacements based on two static CT images acquired at the EE and EI timepoints respectively.

Conclusions: In this study, we have presented an ML-assisted method to estimate voxel displacement based on static CT images. The ML model is expected to be computationally inexpensive during application. We expect this method to serve as an essential step in promoting the feasibility of using regional kinematic biomarkers, such as strain, in diagnostic and prognostic assessment of lung injury patients.

References: [1] Herrmann et al., J Eng Sci Med Diagn Ther. 2022 Nov 1;5(4):040801. [2] Murphy et al. IEEE Trans. Med. Imaging. 30.11 (2011). [3] Neelakantan et al., Proc. SPIE 12928, MI 2024

Acknowledgements (Optional): RA was supported by the NIH, under grant No. R00HL138288.