An Automatic Method for PET Delineation of Cervical Tumors
- Autori: Stefano, A; Vitabile, S; Russo, G; Marletta, F; D'Arrigo, C; D'Urso, D; Gambino, O; Pirrone, R; Ardizzone, E; Gilardi, MC; Ippolito, M
- Anno di pubblicazione: 2015
- Tipologia: Abstract in rivista (Abstract in rivista)
- OA Link: http://hdl.handle.net/10447/211980
Aim: PET imaging is increasingly utilized for radiation treatment planning. Nevertheless, accurate segmentation of PET images is a complex and unresolved problem. Aim of this work is the development of an automatic segmentation method of Biological Target Volume (BTV) in patients with cervical cancer. Materials and methods: Random walks (RW) is a graph-based method that represents a DICOM (Digital Imaging and COmmunications in Medicine) image as a graph. The voxels are its nodes and the edges are defined by a cost function which maps a change in image intensity to edge weights. Then, RW partitions the nodes into target and background subsets. To create an automatic method starting from previous work (A. Stefano, et al. A Graph-Based Method for PET Image Segmentation in Radiotherapy Planning: A Pilot Study, in A. Petrosino, ed., Image Analysis and Processing - ICIAP 2013: LNCS, v. 8157, Springer Berlin Heidelberg, p. 711-720), we propose an automated RW seed localization approach. The algorithm identifies the PET slice with the highest SUVmax and a maximum of 10 target and 8 background seeds for each volume slice. The voxels with a SUV>95% of SUVmax are marked as target seeds. Then, the method explores the hottest voxel neighborhood through searching in 8 directions to identify the background voxels with a SUV<30% of the average SUV of target seeds. Once the target and background seeds are localized, RW performs a 3D lesion delineation. BTV is manually defined by two nuclear medicine physicians in 18 patients with cervical metastases undergoing a 11C-labeled Methionine PET/CT examinations before the radiotherapy treatment. The accuracy of the proposed method is evaluated making a comparison with manual delineation by the dice similarity coefficient (DSC) and median Hausdorff distance (HD). Results: The BTV delineation is not subject to both intra and inter-operator variability. An analysis of the time performance shows that the segmentation time for single slice is around 0.3 seconds. The DSC range of PET delineation is found to be from 81.7% up to 92.6% with a mean of 88.57±3.22%. The HD range is found to be from 1.00mm up to 3.67mm with a mean of 1.96±0.62mm. Conclusion: A slice-by-slice manual PET segmentation is a time expensive method because dozens of slices must be delineated. Results show that our algorithm is automatic and fast, satisfying critical requirements in a clinical environment. In addition, high DSC and low HD values confirm the accuracy of delineation method.