1.
Patients
Our institutional review board approved this study and waived the informed consent requirement.
Chest CT images of 10 patients were obtained by a radiologist after a search of the radiology database.
Each patient underwent CT imaging at baseline and during follow-up (1–12 months apart).
Each scan was obtained at full inspiration using a 16-slice MDCT scanner.
The images were acquired using the following parameters: volume size,
512 × 512 × slices; slice thickness,
5 mm; X- and Y-direction pixel spacing,
0.546–0.644 mm,
and average spacing,
0.60 mm.
2.
Preprocessing
All image data with additional image slices in the gray-level lung volume were interpolated to obtain a resolution closer to isotropic resolution using sub-resample volume programs combining B-spline interpolation.
Finally,
the lung tissue and lung mask volumes in paired temporal chest CT scans from all the patients were segmented.
3.
Image registration
Avants et al.
[1] introduced the greedy symmetric normalization (SyN) method as a lower-cost strategy to implement diffeomorphic transformation for deformable image registration,
which,
by definition,
preserve topology [2,
3].
We performed greedy SyN mapping driven by cross-correlation (CC) and initialized by mutual information (MI) with an affine transformation.
Four-level and five-level image pyramids were used to compute MI and CC,
respectively,
which constituted the MI-CC-SyN implementation [4].
Moreover,
we implemented greedy SyN mapping driven by demons,
initialized by MI with the same affine transformation,
and used the same multiple resolution–optimization parameters as used in MI-CC-SyN implementation,
which constituted the MI-demons-SyN implementation.
The obtained affine and deformation field mapping was used to warp previous images to the corresponding current images with linear interpolations.
Finally,
subtraction images of the lung volume were obtained by subtracting the warped previous volume images from the current volume images.
4.
Accuracy evaluation
The accuracy of image registration implementation was evaluated by an expert using landmark feature points identified in the previous and corresponding current images and warped previous images using a software introduced by Murphy et al.
[5].
The Jacobian of the deformation field is greater than zero for topology-preserving mapping [6].
Jacobian >1 implies local expansion,
Jacobian <1 implies local shrinkage,
and Jacobian =1 implies no change [7].
The log-Jacobian of the deformation field provides a contrast image in which low-contrast areas correspond to local shrinkage,
whereas high-contrast areas correspond to local growing lesions.