MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Two-Stream Deep Learning Method for Ultrasound Image Tracking

H Sun1*, W Zhao2 , L Xing3 , (1) Tsinghua University, Beijing, China (2) Stanford University, Palo Alto, CA, (3) Stanford University School of Medicine, Stanford, CA

Presentations

(Monday, 7/15/2019) 4:30 PM - 6:00 PM

Room: 303

Purpose: Ultrasound image-based tracking is very important for radiotherapy, which ensures the accurate delivery of prescribed dose to the target while sparing the normal tissue during treatment, and improve the safety and quality of treatment. Here, we propose an image tracking method based on deep learning.

Methods: We designed a two-stream deep network to track the target(s) in ultrasound images. The model consists of two parts: the first is the extraction of spatial domain features. We draw on the idea of Mask R-CNN, using Resnet-50 and ROI-Align. Feature extraction is performed on a patch of 40*30 around the target point. The second is the extraction of temporal domain features. In this part, we first used Gunnar Farneback's algorithm to calculate the optical flow between two consecutive frames of pictures and estimated the motion between the two frames. The optical flow variation extraction feature of the surrounding area is represented as a feature of the temporal domain using VGG19. Finally, we performed feature fusion on the two dimensional features to obtain the change of the target area. The network was evaluated using CLUST2015 dataset with 24 training sets and 39 test sets.

Results: The proposed model achieved an accuracy of 1.27±1.09mm in the training set and an accuracy of 1.45±1.27mm in the test set, which is comparable to the state-of-art method today. Furthermore, because of the depth of learning, there is no need to manually extract features. And the proposed method has a lower false detection rate compared with directly using the universal object detection framework.

Conclusion: This study shows that by blending temporal and spatial features, our model can significantly improve the performance of ultrasound image tracking. The use of deep learning models can significantly improve system performance and facilitate the implementation of tasks such as image guided radiation therapy.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email