Evaluation of elastix-based propagated align algorithm for VOI- and voxel-based analysis of longitudinal (18)F-FDG PET/CT data from patients with non-small cell lung cancer (NSCLC)

Abstract

Background: Deformable image registration allows volume of interest (VOI)- and voxel-based analysis of longitudinal changes in fluorodeoxyglucose (FDG) tumor uptake in patients with non-small cell lung cancer (NSCLC). This study evaluates the performance of the elastix toolbox deformable image registration algorithm for VOI and voxel-wise assessment of longitudinal variations in FDG tumor uptake in NSCLC patients.

Methods: Evaluation of the elastix toolbox was performed using (18)F-FDG PET/CT at baseline and after 2 cycles of therapy (follow-up) data in advanced NSCLC patients. The elastix toolbox, an integrated part of the IMALYTICS workstation, was used to apply a CT-based non-linear image registration of follow-up PET/CT data using the baseline PET/CT data as reference. Lesion statistics were compared to assess the impact on therapy response assessment. Next, CT-based deformable image registration was performed anew on the deformed follow-up PET/CT data using the original follow-up PET/CT data as reference, yielding a realigned follow-up PET dataset. Performance was evaluated by determining the correlation coefficient between original and realigned follow-up PET datasets. The intra- and extra-thoracic tumors were automatically delineated on the original PET using a 41% of maximum standardized uptake value (SUVmax) adaptive threshold. Equivalence between reference and realigned images was tested (determining 95% range of the difference) and estimating the percentage of voxel values that fell within that range.

Results: Thirty-nine patients with 191 tumor lesions were included. In 37/39 and 12/39 patients, respectively, thoracic and non-thoracic lesions were evaluable for response assessment. Using the EORTC/SUVmax-based criteria, 5/37 patients had a discordant response of thoracic, and 2/12 a discordant response of non-thoracic lesions between the reference and the realigned image. FDG uptake values of corresponding tumor voxels in the original and realigned reference PET correlated well (R (2)=0.98). Using equivalence testing, 94% of all the voxel values fell within the 95% range of the difference between original and realigned reference PET.

Conclusions: The elastix toolbox impacts lesion statistics and therefore therapy response assessment in a clinically significant way. The elastix toolbox is therefore not applicable in its current form and/or standard settings for PET response evaluation. Further optimization and validation of this technique is necessary prior to clinical implementation.

Keywords: Elastix; Image deformation; NSCLC; PET/CT.

Figure 1

Fusion steps of the propagated align algorithm. Reg-rigid: rigidly registered image. Reg-elastix: rigid and non-rigidly registered image. Step 1: rigid CT to CT alignment of target to reference. Step 2: translation of step 1 on target PET. Step 3: CT to rigid aligned CT non-rigid alignment using elastix toolbox. Step 4: translation of step 3 to PET of step 2. QA: resample of image as quality assurance, so voxel size matches prior to voxel-by-voxel comparison.

Figure 2

Equivalence plot between reference and realigned images of 39 patients. 94% of the voxels are within the 95% range of the difference.

Figure 3

Original (A), elastix adjusted (B), and subtraction(C) maximum intensity projection of patient 6. These images show the typical results before and after alignment as well as the difference between these two images.

Figure 4

Example image of patient 7: baseline (A), reference (B), elastix adjusted (C) and subtraction (D). Because of a mix of non-active and active tumor tissue between the two scans (and no visible decrease in anatomic tumor size with stable disease), only 57% of the voxels aligned within the 95% limits of agreement of −1.46 to 1.46 ΔSUV. However, on crude visual assessment, the images look very similar.

Figure 5

Relation between thoracic imaging alignment and tumor response according to RECIST criteria per patient. RECIST tumor response was not related to the thoracic alignment percentage.