Auto-segmentation of the clinical target volume using a domain-adversarial neural network in patients with gynaecological cancer undergoing postoperative vaginal brachytherapy.

Purpose: For postoperative vaginal brachytherapy (POVBT), the diversity of applicators complicates the creation of a generalized auto-segmentation model, and creating models for each applicator seems difficult due to the large amount of data required. We construct an auto-segmentation model of POVBT using small data via domain-adversarial neural networks (DANNs).

Methods: CT images were obtained postoperatively from 90 patients with gynaecological cancer who underwent vaginal brachytherapy, including 60 and 30 treated with applicators A and X, respectively. A basal model was devised using data from the patients treated with applicator A; next, a DANN model was constructed using these same 60 patients as well as 10 of those treated with applicator X through transfer learning techniques. The remaining 20 patients treated with applicator X comprised the validation set. The model's performance was assessed using objective metrics and manual clinical evaluation.

Results: The DANN model outperformed the basal model on both objective metrics and subjective evaluation (p<0.05 for all). The median DSC and 95HD values were 0.97 and 3.68 mm in the DANN model versus 0.94 and 5.61 mm in the basal model, respectively. Multi-centre subjective evaluation by three clinicians showed that 99%, 98%, and 81% of CT slices contoured by the DANN model were acceptable versus only 73%, 77%, and 57% of those contoured by the basal model. One clinician deemed the DANN model comparable to manual delineation.

Conclusion: DANNs provides a realistic approach for the wide application of automatic segmentation of POVBT and can potentially be used to construct auto-segmentation models from small datasets.