MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Accurate Landmark Pairs Detection for 4DCT Lung Deformable Image Registration Verification

Y Fu*, X Wu , H Li , D Yang , Washington University School of Medicine, St. Louis, MO

Presentations

(Thursday, 7/18/2019) 7:30 AM - 9:30 AM

Room: 225BCD

Purpose: To automatically and precisely detect large quantity of landmark pairs between pairs of intra-patient image volumes for the purpose of deformable image registration (DIR) verification. We expect that the generated landmark pairs will significantly augment the current DIRLAB benchmark datasets in both quantity and positional accuracy.

Methods: Large number of landmark pairs were detected within the lung between the end-exhalation (EE) and end-inhalation (EI) phases of the 10 DIRLAB 4DCT lung datasets. Foerstner operator was applied to find thousands of landmarks within the lung of the EI phase. A parametric image registration method (pTVreg) was used to register the EE and EI phases to establish the corresponding landmarks in EE. Because image registration was subject to registration errors, a Multi-Stream Pseudo-Siamese (MSPS) network was developed to further improve the landmark positional accuracy in EE by directly predicting a vector of xyz shift to optimally align the landmarks in EE to that in EI. Positional accuracy was evaluated using both digital phantoms and random manual spot checks.

Results: Compared to DIRLAB datasets, the quantity of detected landmark pairs increased more than threefold. The mean and standard deviation of target registration error (TRE) was 0.47±0.45 mm with 90% of landmark pairs having a TRE smaller than 1mm for the ten digital phantom cases with known ground truths. Random manual spot checks showed that the proposed procedure performed better than or at least as good as human in detecting landmark pairs.

Conclusion: A novel method was developed to automatically and precisely detect large quantity of landmark pairs between intra-patient volumetric medical image pairs for quantitative evaluation of DIR algorithms.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email