A block matching-based registration algorithm for localization of locally advanced lung tumors

Abstract

Purpose: To implement and evaluate a block matching-based registration (BMR) algorithm for locally advanced lung tumor localization during image-guided radiotherapy.

Methods: Small (1 cm(3)), nonoverlapping image subvolumes ("blocks") were automatically identified on the planning image to cover the tumor surface using a measure of the local intensity gradient. Blocks were independently and automatically registered to the on-treatment image using a rigid transform. To improve speed and robustness, registrations were performed iteratively from coarse to fine image resolution. At each resolution, all block displacements having a near-maximum similarity score were stored. From this list, a single displacement vector for each block was iteratively selected which maximized the consistency of displacement vectors across immediately neighboring blocks. These selected displacements were regularized using a median filter before proceeding to registrations at finer image resolutions. After evaluating all image resolutions, the global rigid transform of the on-treatment image was computed using a Procrustes analysis, providing the couch shift for patient setup correction. This algorithm was evaluated for 18 locally advanced lung cancer patients, each with 4-7 weekly on-treatment computed tomography scans having physician-delineated gross tumor volumes. Volume overlap (VO) and border displacement errors (BDE) were calculated relative to the nominal physician-identified targets to establish residual error after registration.

Results: Implementation of multiresolution registration improved block matching accuracy by 39% compared to registration using only the full resolution images. By also considering multiple potential displacements per block, initial errors were reduced by 65%. Using the final implementation of the BMR algorithm, VO was significantly improved from 77% ± 21% (range: 0%-100%) in the initial bony alignment to 91% ± 8% (range: 56%-100%;p < 0.001). Left-right, anterior-posterior, and superior-inferior systematic BDE were 3.2, 2.4, and 4.4 mm, respectively, with random BDE of 2.4, 2.1, and 2.7 mm. Margins required to include both localization and delineation uncertainties ranged from 5.0 to 11.7 mm, an average of 40% less than required for bony alignment.

Conclusions: BMR is a promising approach for automatic lung tumor localization. Further evaluation is warranted to assess the accuracy and robustness of BMR against other potential localization strategies.

Figure 1

Workflow for block-matching-based registration. (a) Blocks are placed on the reference (planning) image at the surface of the initial gross tumor volume (solid contour). (b) Each block is independently registered to the moving image using an exhaustive search over all translations in a predefined window. The blocks in this center image show the registered location of the blocks from the left image; two additional blocks registered into the current image plane. The weekly physician contour is provided as a dashed line for reference. (c) The resulting transform vectors are aggregated to obtain the global image registration. This corresponds to the couch shift required for patient setup correction.

Figure 2

Block-matching with multiple-candidate registrations. (a) Using an exhaustive search, registrations are obtained that maximize the SCR within the search space of each block. However, the resulting displacement vectors may not adequately represent the “true” registration of the tumor surface. (b) Registration criteria are relaxed such that a single block may have one or more promising registrations (the “candidates”), represented here by multiple displacement vectors originating from the same point. (c) Because each block should only have one “true” registration, a postprocessing step extracts the single most likely displacement per block. (d) Median filtering yields the final displacement vector field, which better captures the registration of the tumor surface. Steps (b)–(d) are performed at each resolution in the pyramid registration scheme.

Figure 3

Mean and standard deviation of the magnitude of block matching errors from registration of artificially deformed images. Results are shown for the initial BMR implementation, MCR, multiresolution pyramid registration with median filtering, and the combined effect of pyramid registration, MCR, and median filtering. The final group (“All”) represents the group mean (G) and random error (σ) for these 12 patients. Differences between BMR algorithms were not statistically significant.

Figure 4

Mean and standard deviation of the magnitude of discrepancies between block-matching-based registration of weekly CT images and deformable mesh registration. Results are shown for the initial BMR implementation, MCR, multiresolution pyramid registration with median filtering, and the combined effect of pyramid registration, MCR, and median filtering. The final group (“All”) represents the group mean (G) and random error (σ) for these 12 patients. Asterisks denote significant error reductions between BMR algorithms.

Figure 5

The effect of the MCR regularization technique on block-matching accuracy for weekly CT images. Beneath each subplot is the patient index (top row) and the fraction of blocks with modified displacements after MCR (bottom row). Registration errors are shown only for this fraction of blocks. The final group (“All”) represents the group mean (G) and random error (σ) for these 12 patients. (a) Comparison of MCR against the initial block-matching implementation. (b) Comparison of MCR against the final stage of pyramid registration. The additional effects of median filtering are also included in both cases.

Figure 6

Minimum detectable BRE with a statistical significance (α) of 0.05 and a power (1−β) of 80%. Results are shown for the initial BMR implementation, MCR, multiresolution pyramid registration with median filtering, and the combined effect of pyramid registration, MCR, and median filtering.

Figure 7

Cumulative histograms of the volume overlap between planned and treatment GTV after bony anatomy alignment and block-matching-based registration. Results are shown for patients stratified according to the presence of PAC such as atelectasis and pleural effusion.