Segmentation of Pelvic Structures for Gynecologic Brachytherapy

Jan Egger, Neha Agrawal, Tyler Blevins, Nabgha Farhat, Guillaume Pernelle, Xiaojun Chen, Yi Gao, Wells, William, Tobias Penzkofer, Tina Kapur, Akila N. Viswanathan · 2012

Methods – In this study we used six T2-weighted magnetic resonance imaging (MRI) datasets from a Siemens 3T scanner. A physician carefully manually segmented the tumor, the bladder, and the rectosigmoid in each dataset for reference. We used the GrowCut [2] algorithm in 3D Slicer [3] which is an interactive segmentation algorithm based on the idea of cellular automata to segment each of the structures. In each case, the initialization of GrowCut was performed on sagittal, axial, and coronal cross-sections. In this initialization step, parts of the structure to be segmented and parts of the background are marked on the image with the Slicer brush tool (a typical initialization for a bladder is shown in three upper images of Figure 1). The algorithm then automatically computed the contours for the structure. Results – Segmentation of the bladder was successfully performed in all cases, with a Dice Score of 91.94±5.4 compared to expert manual segmentation (second row of Figure 1 shows GrowCut, and third row shows manual refinement segmentation). However, GrowCut was not able to achieve satisfactory segmentation results for the tumor and the rectosigmoid. Figure 2 shows the results of GrowCut segmentation of the bladder rendered together with the tumor and sigmoid from manual segmentation. This is attributable to the heterogeneity in appearance of these structures. Conclusions – In this contribution, we studied the segmentation of pelvic structures to support the process of automated contouring for gynecologic brachytherapy. Contouring of the bladder was achieved accurately using the GrowCut algorithm in 3D Slicer. However, manual contouring was needed to achieve segmentation results for the tumor and the rectosigmoid. Future work will include the application of additional methods from the literature for these structures [4, 5].

Read the paper · More papers on PaperTik