Session Item

Automatic organ-at-risk segmentation (OAR) has high potential; ideadly its precission should be comparable to that of clinical experts. The purpose of this study was to train both an open source and a commercial automatic segmentation method for 16 OAR (shown in Figure 1A) according to the Danish Head and Neck Cancer Study Group (DAHANCA) guidelines and to validate both algorithms by comparison with inter-observer variability by clinical DAHANCA experts.

CT scans and clinical delineations from 600 patients from six centres in the DAHANCA 19 randomized study were included. This data was randomly selected into a validation set (N=70) and training batches of increasing (total N=530). This abstract presents the first results of the training set of 50 patients. An experienced oncologist curated training data manually to ensure adherence to DAHANCA guidelines. The final test of the automatic segmentation algorithm was done (only once) in a test set (N=26) with multiple independent delineations by clinical experts from all the DAHANCA centres (median 9 observers per patient). Two convolutional networks were trained: A 3D full-resolution network using the nnU-net open source framework (nnU-net) and a U-net-like network in collaboration with MIM software (MIM). When both networks were considered final by evaluation in the validation set, they were evaluated in the test set using multiple metrics, including the mean surface distance (MSD). For each segmentation under evaluation (both manual and automatic), metrics were calculated pairwise to all remaining manual segmentations for that organ and patient (see Figure 1 B). Following this, the median of these metrics was assigned to each segmentation. The Mann-Whitney U test evaluated differences in metrics between manual and automatic segmentations.

Both networks performed well and segmented all OAR at risk with a precision comparable to clinical experts. The median MSDs are shown in figure 2. For most organs, there was no statistically significant difference in the precision of the experts and the automatic segmentations. Both networks were significantly worse for the PCM_Up and the LarynxSG than the experts but significantly better for the PCM_low and Oral Cavity. In addition, the nnU-net was better than experts in the delineation of the Lips. The oesophagus is not shown in the figures; both algorithms segmented its entire length, whereas the experts only segmented the oesophagus in the head and neck region. As a result, the MSD differences were substantial even though they were very similar in the region relevant to HN cancer radiotherapy.

Even though both networks were trained on only 50 patients, they segmented all organs with a precision similar to clinical experts. Following publication, both networks will be made publicly available, the nnU-net network will be freely available for all to download, and MIM will distribute to the customers that are eligible to Protege AI after regulatory approval.