DNetUnet: a semi-supervised CNN of medical image segmentation for super-computing AI service
- PDF / 2,402,890 Bytes
- 22 Pages / 439.37 x 666.142 pts Page_size
- 54 Downloads / 228 Views
DNetUnet: a semi‑supervised CNN of medical image segmentation for super‑computing AI service Kuo‑Kun Tseng1 · Ran Zhang1 · Chien‑Ming Chen2 · Mohammad Mehedi Hassan3
© Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Deep learning approaches have achieved good performance in segmenting medical images. In this paper, we propose a new convolutional neural network architecture named DNetUnet, which combines U-Nets with different down-sampling levels and a new dense block as feature extractor. In addition, DNetUnet is a semi-supervised learning method, which can be used not only to obtain expert knowledge from the labelled corpus, but also to enhance the performance of learning algorithm generalization ability from unlabelled data. Further, we integrate distillation technique to improve the performance on mobile platform. The experimental results demonstrate that the proposed segmentation model yields superior performance over competition. Since the processing of large medical images and distillation technology is enforced, a supercomputing AI training server is a preference for its application. Keywords Medical image segmentation · Semi-supervised deep learning · U-Net · Distillation
1 Introduction In recent years, medical image segmentation technology based on deep learning has developed rapidly and has the advantage of achieving robust segmentation more effectively. However, most of the deep learning algorithms currently only focus on the study of accuracy or speed, very few studies that pay attention to both accuracy and speed at the same time. Meanwhile, as the matures of deep learning
* Mohammad Mehedi Hassan [email protected] 1
School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen, China
2
Shandong University of Science and Technology, Shandong, China
3
Department of Information Systems, College of Computer and Information Sciences, King Saud University, Riyadh 11543, Saudi Arabia
13
Vol.:(0123456789)
K.-K. Tseng et al.
segmentation algorithm, if it can be extended to a mobile environment, how to design acceptable accuracy and speed would be an interesting research challenge. Therefore, the aim of this research is to design a deep learning architecture to solve major problems in medical image segmentation. This research is conducted on multiple cardiac tissues images, using data from the 2017 Automatic Cardiac Diagnosis Challenge (ACDC). Major contributions of this research include: 1. With existing mainstream networks of medical image segmentation, noticeable information is lost during the multiple down-sampling process, but reduction in the subsequent up-sampling is limited. Therefore, we propose an improved neural network DNetUnet, with a dense block as feature extractor to preserve relevant features for image segmentation. 2. After understanding the advantages and limitations of these two most common loss functions in medical image segmentation: dice loss and cross entropy loss function, we introduce focal loss to deal with class
Data Loading...
