Uncertainty-guided multitask semi-supervised network for the multi-organ segmentation of breast cancer CBCT

The delineation of organs at risk (OARs) and clinical target volume (CTV) is an important step in adaptive radiotherapy (ART). Cone-beam computed tomography (CBCT) images are easy to obtain in radiotherapy (RT).

This study aims to develop an effective CBCT-guided delineation method for breast cancer ART.

A total of 60 planning CT images and 330 CBCT images from 60 patients with breast cancer who underwent breast-conserving surgery were used to develop uncertainty-guided multitask semi-supervised network (UGMNet), which is guided by model segmentation uncertainty and uses intra-task consistency learning to effectively utilize unlabeled data. Branch networks were added to take advantage of the region-level shape constraints and bounding-level distance information of the data. UGMNet was trained with 20% labeled data and a large amount of unlabeled data to obtain the generalized segmentation model (Ours-G). The planning CT image of each patient was then inputted to the generalized model for fine-tuning to establish a personal segmentation model (Ours-P). We compared Ours-G and Ours-P with three classical semi-supervised methods, uncertainty aware mean teacher (UA-MT), dual-task consistency network (DTC), and mutual consistency network (MC-Net+).

Compared with other semi-supervised segmentation model, our proposed method achieved better or equivalent segmentation performance under the same backbone network (3D VNet) and task setting. For CTV delineation, the mean Dice similarity coefficient (DSC) of UA-MT, DTC, MC-Net+, Ours-G, and Ours-P were 0.84, 0.80, 0.60, 0.84, and 0.87, respectively. For the heart, the mean DSC values were 0.82, 0.85, 0.72, 0.86, and 0.89, respectively. For the left lung, the mean DSC values were 0.92, 0.93, 0.91, 0.94, and 0.92, respectively. For the right lung, the mean DSC values were 0.96, 0.94, 0.93, 0.97, and 0.96, respectively. For the spinal cord, the mean DSC values were 0.73, 0.72, 0.77, 0.80, and 0.80, respectively.

The proposed method realizes effective delineation for CBCT-guided ART using a small amount of labeled data and improves the segmentation accuracy of CTV and OARs on CBCT images using personalized modeling.