Regional Estimation of Four-Dimensional (4D) Dynamic Lung Volumes Using Free-Breathing Sagittal MRI
Embedded Files
Abstract:
Patients with thoracic insufficiency syndrome (TIS) present complex breathing mechanics that traditional global Pulmonary Function Testing (PFT) and radiation-intensive CT cannot adequately characterize. We propose a novel, fully automated 4D framework for regional lung volume estimation using non-contrast, free-breathing sagittal MRI. The core novelty is a multi-model deep learning architecture that synergizes you only look once (YOLO) for robust lung area detection with segment anything model (SAM2) for high-precision temporal segmentation, achieving an intersection over union (IoU) and dice similarity coefficient (DSC) above 0.96 and 0.98, respectively. This framework generates independent 3D respiratory signals for each sagittal location, which are then decomposed using Fourier transforms and empirical mode decomposition (EMD) to isolate tidal volume (TV) and functional residual capacity (FRC). Validation on 4D cine dMRI data demonstrates near-perfect linear correlations (r=1.00, p<0.0001) for static volumes across analytical solutions. By enabling non-invasive, radiation-free, and asymmetric assessment of lung function, this multi-model approach provides a granular tool for surgical planning and longitudinal monitoring of pediatric and adult thoracic deformities.
Workflow of analysis respiratory signal: (A) The multi 3D respiration signal generative, (B) Solution 1: Volume based on difference end expiration phase and end of inspiration phase, (C) FFT analysis of the frequency cross sagittal location before remove static signal and after remove static signal, (D) Solution 2: Fourier transform and bandpass filter, (E) Empirical mode decomposition (EMD), (F) The original 3D respiratory signal in green is decomposed into intrinsic mode functions (IMFs) with EMD in blue and Bandpass Filter in red separating near-zero components to obtain filtered respiratory signals for static and dynamic analysis: (F1) dMRI acquisition address at 16 mm from the left boundary- algorithm begins catching the Lung area, (F2) dMRI acquisition address at 64mm from the left boundary-lung area sustainable, the result record the maximum of lung's volume in regional, (F3) dMRI acquisition address at 88mm from the left boundary-dynamic lung area is affected by the dynamic left ventricle heart caused noise spike in green respiration signal as green, the red line remove noise signal reconstruction the lung ideal capability at regional, and blue line EDM algorithm decomposition signal in leverage bandwidth the post-processing signal show contribute and affect of an organ into the respiration signal.
Google Sites
Report abuse