Automatic clinical target volume delineation for cervical cancer in CT images using deep learning.

PURPOSE Accurately delineating clinical target volumes (CTV) is essential for completing radiotherapy plans but is time-consuming, labor-intensive, and prone to inter-observer variation. Automating CTV delineation has the benefits of both speeding up contouring process and improving the quality of contours. Recently, auto-segmentation approaches based on deep learning have achieved some improvements. However, unlike organ segmentation, the CTV contains potential tumor spread tissues or subclinical disease tissues, resulting in poorly defined margin interface and irregular shape. It is not reasonable to directly apply the deep learning segmentation algorithms to CTV tasks without considering the unique characteristics of shape and margin. In this work, we propose a novel automatic CTV delineation algorithm based on deep learning addressing the unique shape and margin challenges. METHODS Our deep learning method, called RA-CTVNet, segments the CTV from cervical cancer CT images. RA-CTVNet denotes our automatic CTV delineation algorithm based on deep learning with Area-aware reweight strategy and Recursive refinement strategy. (1) In order to process the whole-volume CT images and delineate all CTVs in one shot, our method is built upon the popular 3D Unet architecture. We further extend it with robust residual learning and squeeze-and-excitation blocks for better feature representation. (2) We propose area-aware reweight strategy which assigns different weights for different slices. The core is adjusting model's attention to each slice. (3) In terms of the trade-off between providing performance improvements and meeting the limitations of GPU memory, we exploit a new recursive refinement strategy to address margin challenge. RESULTS This retrospective study included 462 patients diagnosed with cervical cancer who received radiotherapy from June 2017 to May 2019. Extensive experiments were conducted to evaluate performance of RA-CTVNet. First, compared to different network architectures, RA-CTVNet achieved improvements in Dice similarity coefficient (DSC). Second, we conducted ablation study. The results showed that compared to the backbone, area-aware reweight strategy increased DSC by 3.3% on average and recursive refinement strategy further increased DSC by 1.6% on average. Then, we compared our method with three human experts. Our RA-CTVNet performed better than two experts while comparably to the third expert. Finally, a multicenter evaluation was conducted to verify the accuracy and generalizability. CONCLUSIONS Our findings show that deep learning is able to offer an efficient framework for automatic CTV delineation. The tailored RA-CTVNet can improve the quality of CTV contours, which has great potential for reducing the burden of experts and increasing the accuracy of delineation. In the future, if with more training data, further improvements are possible, bringing this approach closer to real clinical practice.