Session Item

Contouring of organs-at-risk (OARs) and target volumes is an essential step in interventional radiotherapy (brachytherapy, IRT, BT) treatment planning. However, the delineation quality and time spent on contouring depend on the experience of the radiation oncologist. In the last decades, auto-segmentation algorithms have been developed, including atlas-based methods and deep-learning algorithms based. These methods have the potential to reduce inter-and intra-observer variability and speed up the contouring process. This study evaluated existing standard quantitative geometric measures using atlas-based and deep-learning auto-segmentation methods for OARs and target volume in the pelvis.

MIM v. 7.1.5 (MIM Software Inc., Cleveland, OH), installed on a workstation with Intel Xeon 2274 CPU and 16 GB RAM, was used. A total of 23 patients with cervical cancer who underwent Magnetic resonance (MR-based – 1st fraction) and computed tomography (CT‐based –1st, 2nd,3rd,4th fractions) interventional radiotherapy (IRT) were included in this study. All the enrolled patients received intracavitary high‐dose‐rate IRT. Planning CT data were acquired on Optima CT 580 (GE, General Electric) system set on helical scan mode. CT images were reconstructed using a matrix size of 512 × 512 and thickness of 0.625 mm. The planning CT volumes of involved patients for the first and second IRT treatment fraction were collected. Rectum, bladder, small bowel and target volumes (high (HR) and intermediate risk (IR) clinical target volume (CTV) were manually contoured on each CT. An ad hoc workflow was optimized in MIM in order to perform a rigid registration followed by a deformable registration and the subsequent automatic creation of the region of interests (ROIs) on the second CT. The manual ROIs were therefore compared to the automatic ROIs with the use of the Dice Similarity Coefficient (DSC: (2│A∩B│/│A│+ │B│) and the Jaccard Similarity Coefficient (JSC: (│A∩B│/ │AUB│). It is generally accepted that avalue of DSC and JSC > 0.7 represents excellent agreement.

Among all structures, the best results were obtained for bladder segmentation with median DSC and JI values of 82% and 70%, respectively. It is mainly because the bladder has a relatively regular shape and clear boundary in the planning CT images. Automatic segmentation also achieved a good result for HR-CTV (median DSC 79% and median JSC 66%) and IR-CTV (median DSC 88% and median JSC 79%) clinical target volumes.

The most inferior segmentation accuracies were observed on the segmentations of rectum and small bowel (DSC = 64%, JI = 48% and DSC = 52%, JI = 35%, respectively).

We presented a deep learning‐based method using MIM v.7.1.5 architecture to automatically segment the target volumes and OARs in the planning CT images for cervical cancer IRT. Quantitative evaluation results showed that the proposed method could segment the HR/IR‐CTV and bladder with relatively good accuracy.