A Mask R-CNN Model with Improved Region Proposal Network for Medical Ultrasound Image
In medical ultrasound image processing, it is often necessary to select the ROI before segmentation to obtain better segmentation accuracy. With the development of deep learning, the technology of object detection can well implement the function of automatically selecting ROI. The combination of object detection and image segmentation has also been proposed, such as Mask R-CNN, an end-to-end image segmentation model. However, the ROI selection by the algorithm above cannot meet the needs of medical image segmentation. Because its RPN layer is inherited from Faster R-CNN, a target classification framework. What we need is a region that can cover the whole object area with the details of edge. This information has an important influence for the further segmentation. Therefore, this paper improves the selection criteria of the anchor in the RPN layer, making the improved RPN layer more suitable for image segmentation tasks. Finally, the experimental results show that the improved model can achieve higher segmentation accuracy with the appropriate parameters selected.
KeywordsImage segmentation Deep learning Mask R-CNN Ultrasound image Machine learning
This work was supported by the National Natural Science Foundation of China (Grant No.31201121, No.61373109 and No.61403287), the Natural Science Foundation of Hubei Province (Grant No.2014CFB288) and Open foundation of Hubei Province Key Laboratory of Intelligent Information Processing and Real-time Industrial System (Grant Nos.ZNSS2013A0001 and ZNSS2013A004).
- 1.Samundeeswari, E.S., Saranya, P.K., Manavalan, R.: Segmentation of breast ultrasound image using regularized K-means (ReKM) clustering. In: International Conference on Wireless Communications, Signal Processing and NETWORKING, pp. 1379–1383. IEEE (2016)Google Scholar
- 2.Li, L., Lin, J., Li, D., Wang, T.: Segmentation of medical ultrasound image based on Markov random field. In: The International Conference on Bioinformatics and Biomedical Engineering, pp. 968–971. IEEE Xplore (2007)Google Scholar
- 3.Nguyen, T.D., Sang, H.K., Kim, N.C.: Surface extraction using SVM-based texture classification for 3D fetal ultrasound imaging. In: International Conference on Communications and Electronics, pp. 285–290. IEEE (2007)Google Scholar
- 4.He, K., Gkioxari, G., Dollár, P., Girshick, R.: Mask R-CNN (2017)Google Scholar
- 5.Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: towards real-time object detection with region proposal networks. In: International Conference on Neural Information Processing Systems, vol. 39, pp. 91–99. MIT Press (2015)Google Scholar
- 6.Long, J., Shelhamer, E., Darrell, T.: Fully convolutional networks for semantic segmentation. In: Computer Vision and Pattern Recognition, vol. 79, pp. 3431–3440. IEEE (2015)Google Scholar
- 7.Xie, S., Girshick, R., Dollar, P., Tu, Z., He, K.: Aggregated residual transformations for deep neural networks, pp. 5987–5995 (2016)Google Scholar
- 8.Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection, pp. 936–944 (2016)Google Scholar
- 10.Girshick, R.: Fast R-CNN. Computer Science (2015)Google Scholar
- 12.He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition, pp. 770–778 (2015)Google Scholar
- 15.Andrews, S., Hamarneh, G.: Multi-region probabilistic dice similarity coefficient using the Aitchison distance and bipartite graph matching. Computer Science (2015)Google Scholar