Example of PET imaging for response assessment. Early biological changes can be detected by comparing the pretreatment (left) with mid or post treatment (middle). However, PET images suffer from significant levels of noise and poor resolution, and a simple difference comparison (right panel) is not efficient to automatically characterize biological changes. The difference image (right) shows voxels with tumor regression or signal enhancement. We propose advanced tools to extract biological changes from noise to classify the image into responsive and nonresponsive regions based on spatial clustering.
Level set clustering for automated response assessment for clinical studies. Inputs, shown as white rectangles, are pretreatment and posttreatment PET scans, planning dose, and structures. The algorithm steps, shown as yellow rectangles, correct for anatomical changes using a deformable registration, then create a map of biological changes in pre and post PET scans that are further processed to identify patterns using an advanced clustering method mark of response or progression. This information is correlated with planning dose and contours for a detailed report of treatment outcomes, as well as possible treatment errors and their locations.
Accuracy of deformable registration. The pretreatment dataset, used as fixed image for the registration algorithm is shown in (a). The post-treatment dataset is shown in (b), with arrows marking regions of mismatch. Results of applying a deformable registration is shown in (c), where the previous discrepancies have been corrected by the registration algorithm. Effects of applying the registration algorithm on the PET dataset is shown in the next three panels, with the pretreatment PET dataset (d), post-treatment (e) and post treatment PET warped with the deformable registration algorithm (f), shown as overlays on the pretreatment CT dataset. One notes the significantly different location of the remaining SUV activity in the post-treatment dataset when the deformable registration is used. The remaining activity also better matches the underlying anatomy.
Comparison of standard thresholding with clustering algorithms. In the left panel, a coronal slice of the original difference image is shown with intensities levels documented by the scalar bar. When applying standard image analysis algorithms (middle panel) to detect enhancing regions, results are corrupted by noise and over segmentations. This is illustrated in the middle panel by the 3D contour marking detected regions. The level-set based algorithm (right panel) correctly identifies only regions near the original tumor site with enhanced signal.
Evolution of the level set clustering algorithm. Algorithm output at various iterations is shown as a 3D wireframe overlaid on a coronal slice through the difference dataset. As the algorithm morphs initial estimation to regions of signal enhancement, approach advantages are seen at iteration 200 when the shape splits to form multiple regions. Final shape highly conforms to regions of signal enhancement. No specific cluster shape is assumed a- priori by the algorithm.
Algorithm advantage in using spatial information to discern noise from enhancements. In this clinical case, main tumor activity in the pretreatment scan (left) disappears in the post-treatment scan (middle) but there is an enhancing area, barely enhancing at arrow, with borderline intensities of 2–3 SUV. This area delimited by a black contour in the difference image (right) is automatically detected by the algorithm and is marked as suspicious. Maximum SUV decreased from 21.5 to 6.8 SUV, and this case would be classified as responsive if using only simple statistical measures.
Upper row shows a typical clinical case with clusters enhancing (#1, #2), remaining fairly unchanged (#3, #4), and signal decline (#5). In the result panel, enhancing clusters detected by the algorithm are shown as a black contour superimposed on the difference image. Lower row shows a different case in which the algorithm was instrumental in detecting enhancing voxels near the tumor bed. For consistency, the same range of SUV levels were used to display pre and post PET scans.
Example of geometric miss detected by the clustering procedure. The post-treatment PET scan displayed an enhancing cluster at the lower neck level (b). At this location, the SUV activity has borderline background levels, barely discernible when using standard window levels (a). Location borders the treatment field receiving lower dose during treatment (c), thus the region was classified as a treatment error due to inaccurate targeting.
Treatment failure due to different head rest frame between two scans. (a) The simulation CT used in radiotherapy planning is shown with a headrest frame used during scanning. A PET dataset for the same patient was acquired before treatment, to better localize the tumor. As shown in (b), the scanner tabletop had a slightly bent headrest frame, causing the patient to lift his chin up relative to his body. This change is seen in (c) where the external contour segmented from the CT dataset shown in (a) is overlaid on the PET scan shown in (b). To create a treatment plan, the PET dataset was fused to the planning CT using a standard rigid registration that ignored chin movement. The rigidly fused PET dataset is shown in (d), with the contour line representing the 3.5 SUV contour. However, the true tumor location is shown in (e) as a black outline, where a deformable registration was used to take into account changes in chin position. The area inside the common contour in (d) and (e) was treated, while the tumor was actually situated in inside the black contour. Reviewing the post-treatment scan in (f), the tumor progressed in the area not included in the treatment field, shown with arrow.
Article metrics loading...
Full text loading...