Program Information
Automatic Delineation Strategy for Brain Metastases Using Deep Convolutional Neural Network
Y Liu1*, S Stojadinovic2 , B Hrycushko2 , Z Wardak2 , W Lu2 , Y Yan2 , S Jiang2 , X Zhen3 , R Timmerman2 , R Abdulrahman2 , L Nedzi2 , X Gu2 , (1) Sichuan University, Chengdu, ,(2) UT Southwestern Medical Center, Dallas, TX, (3) Southern Medical University, Guangzhou, Guangdong
Presentations
SU-K-FS4-6 (Sunday, July 30, 2017) 4:00 PM - 6:00 PM Room: Four Seasons 4
Purpose: Stereotactic radiosurgery is commonly used in treating small brain lesions and its treatment quality heavily relies on accurate target delineation. The aim of this study is to develop deep convolutional neural network based auto-segmentation strategy to accurately and efficiently delineate small targets on contrast-enhance T1-weighted (T1c) Magnetic Resonance Images (MRI), the standard treatment planning modality.
Methods: The developed deep convolutional neural network architecture for automatic delineation consists of four sections. The first input section extracts concentric image patches with different resolutions and feeds them to convolution section. The second convolution section is made up of three sub paths, where each sub path is a stack of convolutional filters, to capture multi-scale information. The third section is a group of full connected neurons to fuse the features. The fourth classification section finishes voxel-wise prediction to generate a segmentation map. The developed auto-segmentation has been validated on the brain metastases patients dataset with physician manual draw contours as ground truth and the average size of lesions is 0.439±0.840cc.
Results: Evaluation results show the strategy is capable of maintaining the average DICE coefficient (DCs) of 0.67±0.03, mean value of surface-to-surface distance (MSSD) of 0.96±0.30mm and standard derivation of surface-to-surface distance (SDSSD) of 0.77±0.20mm. The area under ROC curve of our segmentation system is 0.98. Compared with our previous intensity-based automatic segmentation strategy, the proposed method can capture the small lesions (<1.5cc) robustly within 2 minutes.
Conclusion: The proposed delineation strategy could automate delineation of different sizes lesions in SRS. It will be useful in clinical practice and helpful in arising the efficiency and accuracy of SRS treatment planning.
Contact Email: