RT Journal Article SR Electronic T1 Lesion detection with extremal regions in thoracic PET-CT images JF Journal of Nuclear Medicine JO J Nucl Med FD Society of Nuclear Medicine SP 1783 OP 1783 VO 56 IS supplement 3 A1 Yang Song A1 Weidong Cai A1 Yun Zhou A1 Dagan Feng A1 Michael Fulham YR 2015 UL http://jnm.snmjournals.org/content/56/supplement_3/1783.abstract AB 1783 Objectives To develop an automated method to detect lesions in thoracic FDG PET-CT images.Methods Our method used extremal region generation and thresholding. First, a nested hierarchy of extremal regions was generated using the maximally stable extremal regions algorithm for each axial slice, based on the SUV data that were linearly rescaled to grayscale. Each extremal region represented an area of homogeneous SUVs. A 2D lesion area was typically covered by one or several extremal regions hence the total number of regions was small. Next, extremal regions with average SUVs higher than a threshold were defined as lesions. The threshold was computed as the average SUV of the middle volume of the image (roughly representing the mediastinum) plus 15% of the maximum SUV. The threshold definition was similar to what was done in our previous work [1] but simplified by removing the mean-shift clustering. 3D connected component analysis was performed to obtain lesion objects.Results The method was tested on 32 NSCLC FDG PET-CT studies that had 61 lesions. A detected lesion object with at least 50% overlap with the ground truth annotation was considered as true positive. Our method obtained 95.1% recall and 96.7% precision for lesion detection. Compared to the standard thresholds SUV-2.5 and 50% SUVmax, our method provided a 33% and 20% improvement in precision; there was no change in recall compared to SUV-2.5 but a 13% improvement for 50% SUVmax. Compared to the more complicated methods of mean-shift clustering with SUV threshold [1] and further refinement with graph cut [2], our method had a 12% and 6% improvement in precision; there was minimal change in recall with a 3% improvement over mean-shift clustering and no change over graph-cut approaches.Conclusions Our extremal region-based method offered improved performance in the detection of lesions in thoracic PET-CT scans and more simplified implementation than clustering and graph cut approaches.Research Support ARC grants.