Premium
Poster — Wed Eve—44: CO‐Registered Multi‐Modality Pattern Analysis Segmentation System (COMPASS) for Radiation Targeting of Head and Neck Cancer Using FDG PET/CT
Author(s) -
Yu H,
Caldwell C,
Mah K,
Poon I,
Balogh J,
MacKenzie R
Publication year - 2009
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.3244148
Subject(s) - voxel , segmentation , artificial intelligence , nuclear medicine , medicine , head and neck cancer , head and neck , radiation therapy , radiation treatment planning , medical imaging , computer science , radiology , pattern recognition (psychology) , surgery
Previous attempts to segment tumours for radiation therapy targeting based on FDG‐PET image thresholds have had little success. However, if the texture information available in PET and CT images is used, more accurate and reliable differentiation of abnormal and normal tissues may be possible. Objective: To develop an automated image segmentation method for head and neck cancer (HNC) using texture analysis of co‐registered FDG‐PET/CT images. Methods: CO‐registered Multi‐modality Pattern Analysis Segmentation System (COMPASS) was developed using a region‐of‐interest‐based Decision Tree K‐Nearest‐Neighbors (DTKNN) classifier. 14 PET and 13 CT texture features such as coarseness, busyness and Left/right symmetrical ratio were calculated for each voxel from corresponding PET and CT images within a window centered on the voxel. Then the voxel was classified as “tumor” or “non‐tumor” using the DTKNN classifier. PET/CT images of 10 patients with HNC who had their primary tumors and positive nodes manually segmented by three radiation oncologists were used for evaluation. Results: The sensitivity per patient was 83%±19% when “true positive voxels” were defined as those voxels identified by at least two physicians as tumor. The specificity was 95%±2% when “true negative” voxels were all soft tissue voxels not identified by any of three physicians as tumor. Results of COMPASS were significantly better than those of three previously published PET threshold‐based methods. Conclusions: This work suggests that an automated segmentation method based on texture classification of FDG‐PET/CT images has the potential to provide accurate delineation of HNC.