Encrypted login | home

Program Information

Automatic Segmentation of Clinical Target Volume and Organs at Risk in Planning CT of Rectal Cancer with Deep Dilated Convolutional Neural Networks

no image available
K Men

K Men, J Dai* ,National Cancer Centre/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing,China


MO-F-205-3 (Monday, July 31, 2017) 4:30 PM - 6:00 PM Room: 205

Purpose: Delineation of clinical target volume (CTV) and organs at risk (OARs) for radiotherapy is time-consuming and prone to inter-observer variation. We firstly proposed a Deep Dilated Convolutional Neural Networks (DDCNN) based method for fast and consistent auto-segmentation of these structures.

Methods: DDCNN is popular in computer vision community. It is an end-to-end architecture enabling fast training and testing. In DDCNN, we firstly deployed dilated convolutions to extract multi-scale context information at early layers, which contains fine boundary information and is very useful for accurate auto-segmentation. Then, we enlarged the receptive fields of dilated convolutions at the end of networks for capturing complementary context information. We used the data from 278 patients with rectal cancer for evaluation. CTV and OARs were delineated and validated by a senior radiation oncologist in planning CT images. The randomly chosen 218 patients were used for training and the rest 60 for validation. The Dice Similarity Coefficient (DSC) was used to measure the segmentation accuracy.

Results: The performance of our proposed method was evaluated on segmentation of CTV and OARs. The average DSC values were 0.877 for CTV, 0.934 for bladder, 0.921 for left femoral head, 0.923 for right femoral head, 0.653 for intestine and 0.618 for colon. And the test time was 45 seconds per patient for all the CTV, bladder, left and right femoral head, colon and intestine segmentation. We analyzed the approaches and results in the literature and observed that our system showed superior performance and faster speed.

Conclusion: We concluded that DDCNN can segment the CTV and OARs accurately and efficiently.

Funding Support, Disclosures, and Conflict of Interest: This work was supported by the National Natural Science Foundation of China [grant numbers 11605291] and the National Key Projects of Research and Development of China [grant number 2016YFC0904600]. The authors report no conflicts of interest with this study.

Contact Email: