Program Information
A Semi-Automatic Algorithm for Segmenting Cervical Tumors in 3D 18FDG PET
L Chen*, G Maquilan , K Thomas , C Shen , Z Zhou , M Folkert , K Albuquerque , J Wang , UT Southwestern Medical Center, Dallas, TX
Presentations
WE-G-201-1 (Wednesday, August 2, 2017) 4:30 PM - 6:00 PM Room: 201
Purpose: To develop a semi-automatic method based on variational model (VM) for accurate segmentation of cervical tumors in 3D ¹⁸FDG PET images.
Methods: In PET, cervical tumors are often partially connected to bladders with similar intensities. Thus conventional segmentation methods such as automatic thresholding and region growing methods will misclassify bladder as cervical tumor. In the proposed VM-based algorithm, cervical tumor is first contoured by an expert at the central slice. This manually segmented slice is used to guide auto-segmentation of remaining slices, where the VM is to explicitly consider the continuity of tumor shape and location among neighboring slices. Initial contours required by VM are obtained by a graph theoretic model. The approximated segmentation from the graph theoretic model is fine-tuned by the proposed similarity based VM under the guidance of tumor manually segmented at the central slice. The proposed VM algorithm is tested on PET images of 57 cervical cancer patients whose cervix and bladder are connected with similar activity values. Using expert-segmented 3D tumors as reference, the dice similarity coefficients (DSC) are calculated to characterize the segmentation accuracy of different segmentation algorithms.
Results: The average DSC value of segmented cervical tumors from 57 patients by using the proposed VM-based algorithm is 0.83 while the automatic thresholding and region growing methods can only achieve 0.60 and 0.52 DSC values, respectively. The VM-based algorithm consistently achieves higher DSC value.
Conclusion: The proposed VM-based algorithm can provide a more accurate way for segmenting cervical tumors in 3D ¹⁸FDG PET images.
Contact Email: