Session Item

MR images are frequently used to improve the accuracy of target and OAR delineation for RT planning that is currently based on CT images. The latest achievements on synthetic CT solutions enable MR-only RT planning by performing dose calculation using MR images only. This emerging scenario can be further augmented by developing automated OAR delineation for MR images. Since most commercially available auto-contouring tools are CT-based, the goal of this work was to demonstrate the feasibility of deep learning (DL) based head and pelvis OAR delineation in T2w MR images.

The input of the auto-contouring was a standard T2w MR image. For both anatomical sites the DL models were trained and tested using one dataset with 90-10% separation. The head models were developed on a public dataset [DOI:10.7937/tcia.2019.bcfjqfqb] including 55 T2w cases. The number of manually contoured cases per organ varied between 22 and 41 depending on coverage. The pelvis models were developed on a set of 48 T2w prostate cases, where all organs were manually contoured. Both datasets were contoured by medical students trained by radiation oncologists.

The segmentation method was an ensemble of 2D and 3D convolutional neural networks. The organs were segmented in their bounding box using a 3D model, and the center of the bounding box was localized using 2D (axial, coronal, and sagittal) models. If the full coverage of a structure was not guaranteed (body, brain, bowel-bag) only 2D axial model was used. All 2D and 3D models were trained for each organ, separately.

The auto-contours were compared with the manual contours using DICE and Surface DICE (SDICE) metrics. The first measures volumetric overlap, while the second measures how the surface of two objects overlaps within a predefined tolerance. The tolerance was set to 1 mm for head and 2 mm for pelvis organs. For each organ the mean metrics were computed on all test cases. The auto-contours were rated by 2 radiation oncologists using 1-5 score based on their clinical usability and the mean score was computed for each organ.

Table 1 shows the mean accuracy of auto-contours measured on the test cases. Among head organs the best accuracy (DICE>87%, SDICE>94%, Score>3.9) were achieved for body, brainstem, cochlea, eye, lens, and spinal cord, while mandible, PCM, and Larynx G had the lowest performance. Among pelvic organs the best accuracy (DICE>92%, SDICE>87%, Score>4.2) were achieved for body, bladder, and femoral heads, while bowel-bag and urethra had the lowest performance. Figure 1 demonstrate the auto-contours for one representative head and pelvis case.

Table 1. Accuracy of the automated organ delineation (SDICE = Surface DICE)

Figure 1. Representative results of the automated organ delineation

The presented method can provide organ contours which demonstrate accuracy that is good enough for clinical use. Ongoing work is being undertaken on expanded dataset to further increase the accuracy for all head and pelvic organs.