SU-E-J-106: Atlas-Based Segmentation: Evaluation of a Multi-Atlas Approach for Lung Cancer

Previous studies have shown atlas-based segmentation using a single best matched (SBM) atlas subject can significantly reduce contouring time. A new multi-atlas approach has been shown to provide greater accuracy than SBM for cancer of the head and neck. The goal of this study was to evaluate the multi-atlas technique for lung cancer treatment planning.

An institution's SBRT lung atlas containing 82 subjects was utilized for atlas segmentation. Each atlas subject contained manually defined contours of the esophagus, cord, heart, left lung, right lung, and trachea. CT scans and contours for 16 subjects were evaluated. SBM used the one automatically determined best match for segmentation. Multi-atlas used multiple automatically determined best matches: 3, 4, and 5, respectively. The final segmentation for multi-atlas was generated using Majority Vote which comprises the area of overlap for at least half of the individual segmentations (2 of 3, 2 of 4, and 3 of 5, respectively). Average Dice Similarity Coefficients (DSC) were calculated for each structure to compare against manually defined 'gold' standard contours for that subject. Overall percent improvement was calculated as the proportion of the error corrected by the method, or % difference on 1-DSC.

All multi-atlas methods were significantly more accurate than SBM (p-value < 0.0005) with average DSC of 0.802 +/- 0.172, 0.809 +/ 0.163, 0.802 +/- 0.182 respectively for Multi-3, Multi-4, and Multi-5 compared to 0.773 +/- 0.187 for SBM. No significant differences existed between the different multi-atlas approaches. Overall, Multi-4 showed the greatest improvement over SBM with 16% improvement followed by Multi-3 and Multi-5 at 12%.

Each multi-atlas approach resulted in significantly more accurate contours compared to the SBM. While still requiring some editing, this method for segmentation using multiple atlases shows promise for further decreasing the contouring time required for lung cancer. MIM Software Inc.