Uncertainty-aware segmentation of the cervical cancer high-risk clinical target volume from T2-weighted MRI using evidential deep learning

Accurate contouring of the high-risk clinical target volume (HR-CTV) is vital for effective radiation therapy of cervical cancer, particularly in brachytherapy, where precise dose delivery is crucial for successful treatment. Several deep neural network models have been recently developed and demonstrated promising performance in HR-CTV segmentation, but the model performance depends on the quality of the image and contour data used for training, which often significantly varies. Existing models lack mechanisms for estimating the uncertainty level of their inference, thus leading to potential overconfidence in the generated segmentations. This study proposes an uncertainty-aware segmentation method based on evidential deep learning approach which estimates confidence of its inference at the same time of computing HR-CTV segmentation from MRI. We employed a dual convolution-transformer U-Net as a backbone to compute the probability for segmentation for each voxel, and incorporated an uncertainty block to the final output of the network for uncertainty estimation. A total of 250 T2-weighted MRIs obtained from 138 cervical cancer patients were used to train, validate, and test the model. For 27 test cases, four sets of manual segmentations were obtained from three radiation oncologists to assess the segmentation and uncertainty estimation performance. In comparison to the consensus of manual contours, the proposed model achieved dice similarity coefficient (DSC) and Hausdorff distance (HD95) of 0.713±0.106 and 9.309±4.474, respectively. We observed significant inter/intra-rater variability (DSC: 0.656-0.709; HD95: 9.665-12.364), and the proposed model showed comparable performance to the experts’ manual segmentations. Furthermore, the estimated uncertainty represented well the variability in multi-rater segmentations, demonstrating its effectiveness in capturing difficult to segment regions. The experimental results demonstrate that including uncertainty estimates in HRCTV segmentation leads to a more comprehensive understanding of model prediction, and potentially aids clinicians in making more informed decisions and improving treatment outcome.