Voxel clustering for quantifying PET-based treatment response assessment

Authors


Abstract

Purpose:

Imaging biomarkers are crucial in managing treatment options for cancer patients. They are extremely powerful tools since they allow personalized treatment assessment early during therapy by using repeated imaging to detect and quantify tumor response. Currently, treatment response assessment from consecutive imaging is measured by simple global measures that do not capture a tumor's heterogeneous response. The authors present an automated, multivoxel metric that groups voxels into clusters of changes for a local definition of radiation treatment efficiency from multiple PET imaging studies acquired at different time periods for assessing therapeutic response.

Methods:

The algorithm employs level-set mathematics to extract changing features to classify voxels into response patterns. First, pretreatment and post-treatment PET images were aligned using a deformable registration to correct for posture and soft tissue changes. The detailed mapping was modeled by free form deformations B-spline optimized using the limited memory L-BFGS algorithm. The posture-corrected datasets are then subtracted to produce an image of molecular changes embedded with noise. Once images were aligned and subtracted, a segmentation algorithm combining the concepts of voxel and distance-based techniques classified voxels into patterns of signal reduction or enhancement. Although signal reduction is evidence of successful treatment, signal-enhancing regions are an indication of treatment failure. For an in depth analysis of potential treatment errors, patterns of signal enhancement were correlated with the radiation treatment dose and anatomical structures from the treatment plan using image registration methods.

Results:

The algorithm was retrospectively applied to PET/CT and radiotherapy (RT) oncology data from an NCI-sponsored clinical trial (81 clinical cases from RTOG 0522 Trial) for combined drug and radiation therapy in head and neck carcinomas. This clinical trial dataset presented a realistic environment for implementing and validating our algorithm to correlate local response as observed in serial PET with delivered dose. The technique was instrumental in detecting geographical and segmentation misses on the actual clinical cases by providing accurate voxel-by-voxel analysis of metabolic changes. Results of the level-set based clustering algorithm are saved as a detailed report of enhancing/nonenhancing regions and their location, and can be further displayed as a colorwash laid over the original anatomy for in depth analysis.

Conclusions:

The automated technique was instrumental in analyzing treatment response in the clinical cases and provided an useful tool for accurate, outcome-based response assessment of the radiation treatment process. The developed method is general and should be extendable to other high-resolution diagnostic imaging with minor modifications.

Ancillary