Click here to


Are you sure ?

Yes, do it No, cancel

Lung Feature Tracking in 4DMRI Using a Scale-Invariant Feature Transform Method

E Colvill1*, E Gaberdiel2 , A Lomax1 , O Bieri3 , Z Celicanin3 , D Weber1 , G Fattori1 , (1) Paul Scherrer Institute, Villigen, Switzerland, (2) Cambridge University, Cambridge, United Kingdom,(3) Division of Radiological Physics, Department of Radiology, University Hospital Basel, University of Basel, Basel, Switzerland and Department of Biomedical Engineering, University of Basel, Basel, Switzerland


(Sunday, 7/14/2019) 4:30 PM - 5:00 PM

Room: Exhibit Hall | Forum 8

Purpose: The purpose of this study is to demonstrate the feasibility of a scale invariant feature transform (SIFT) method for automated lung motion characterization from 4D-MRI scan data.

Methods: 4D-MRI scans of 20 healthy volunteers and 3 patients were acquired. The data consists of 3D stacked 2D images with slice thickness of 5mm representing the anatomy with a frequency of 3.33 Hz. A SIFT-based method was employed to automatically extract features and corresponding trajectories tracked on the majority (>50%) of volumes with fewer than 20 misses. Misses and outlier data points were replaced by cubic interpolation. Outliers were identified as data points with a distance from the trajectory centroid greater than a threshold, defined as [Threshold = P₇₅ + 1.5 x P₇₅₋₂₅] where P₇₅ is the 75ᵗʰ percentile of the points-to-centroid distance distribution and P₇₅₋₂₅ is the interquartile range (IQR). The locations of the features within the lung and motion trajectory information about each feature were calculated. Motion information included the mean range of the trajectory, IQR and 95ᵗʰ percentiles.

Results: Preliminary results from five volunteer 4DMRI sets show that on average six features were tracked for each set, with the majority of features in the lower-to-middle portion of the lung. The average mean range, IQR and 95ᵗʰ percentiles of the feature trajectories were 4.8 (±1.9)mm, 4.3 (±2.1)mm and 5.2 (±2.0)mm respectively, with inferior features displaying greater motion ranges than superior.

Conclusion: Automatic lung motion characterization based on 3D image stacks built using 4DMRI was successfully demonstrated using a SIFT-based method. The features were extracted from the lung and tracked throughout the 4DMRI series with features in different portions of the lung displaying different motion characteristics.


Feature Extraction, MRI, Lung



Contact Email