Due to the low contrast between tumors and surrounding tissues in CBCT images, this study was designed to propose an automatic segmentation method for central lung cancer in CBCT images.

There are 221 patients with central lung cancer were recruited. Among them, 176 patients underwent CT localization and 45 patients underwent enhanced CT localization. The enhanced CT images were set as the lung window and mediastinal window, and elastic registration was performed with the first CBCT validation images to obtain the paired data set. The CBCT images could be transformed into" enhanced CT" under the lung window and mediastinal window by loading the paired data sets into cycleGAN network for style transformation. Finally, the transformed images were loaded into the UNET-attention network for deep learning of GTV. The results of segmentation were evaluated by Dice similarity coefficient (DSC), Hausdorff distance (HD) and the area under the receiver operating characteristic curve (AUC).

The contrast between tumors and surrounding tissues was significantly improved after style transformation. The DSC value of cycleGAN+ UNET-attention network was 0.78±0.05, HD value was 9.22±3.42 and AUC value was 0.864, respectively.

The cycleGAN+ UNET-attention network can effectively segment central lung cancer in CBCT images.

Deep learning; Style transformation; Target automatic segmentation; Lung cancer