Investigations on Performances of Pre-trained U-Net Models for 2D Ultrasound Kidney Image Segmentation
The importance of segmentation in medical applications is inevitable and hence its performance and accuracy is a priority. The semi-automated and automated segmentation of kidneys from 2D ultrasound images is quite challenging due to the intensity distrib
- PDF / 2,485,722 Bytes
- 11 Pages / 439.37 x 666.142 pts Page_size
- 49 Downloads / 168 Views
and D. Abraham Chandy
Department of Electronics and Communication Engineering, Karunya Institute of Technology and Sciences, Karunya Nagar, Coimbatore 6411114, India [email protected]
Abstract. The importance of segmentation in medical applications is inevitable and hence its performance and accuracy is a priority. The semi-automated and automated segmentation of kidneys from 2D ultrasound images is quite challenging due to the intensity distribution differences within the kidney and also due to the intensity similarity with the nearby organs. Deep learning have paved its way into outperforming traditional techniques in various fields of applications efficiently, but the amount of data used is imperious. The relevance of deep learning in biomedical application is also inevitable as it makes the application completely automatic and precise. This paper investigates the performances of pre-trained U-Net model using various backbones for segmentation of kidneys from 2D ultrasound images. Experimentation results obtained shows that U-Net model with VGG-16 backbone outperformed with a promising accuracy of 0.89, thus demonstrating that segmentation can be done even with limited count of images within the dataset. Keywords: Kidney segmentation · Ultrasound · Deep learning · U-Net · Data augmentation · Transfer learning
1 Introduction In medical field, the two main cognitive and challenging tasks faced by diagnostic experts are the analysis and interpretation of the acquired image from any medical imaging modality. Over the years, researchers have come up with various semi-automated and automated techniques and technologies to provide solutions for the same. Most recent techniques presented by researchers involves deep learning for these applications. The most competent wheels within wheels machine learning category that as paved its way in various applications such as biometrics, speech recognition, object detection etc. [1] is Deep Learning. The results obtained for deep learning techniques is promising yet it faces few challenges [2]. One of the many challenges is the need for a large dataset to obtain computational efficiency. When considering biomedical field, especially medical image analysis, the procurement of a large dataset is quite demanding [3]. © ICST Institute for Computer Sciences, Social Informatics and Telecommunications Engineering 2020 Published by Springer Nature Switzerland AG 2020. All Rights Reserved M. H. Miraz et al. (Eds.): iCETiC 2020, LNICST 332, pp. 185–195, 2020. https://doi.org/10.1007/978-3-030-60036-5_13
186
D. M. Alex and D. Abraham Chandy
Inorder to overcome this problem, the researches presented data augmentation [4] has a pre-processing technique. Data augmentation is described as a method used to escalate the dataset content by applying various techniques such as cropping, padding, flipping etc. on the images present in the dataset [5]. Another way of applying deep learning on small dataset is by using transfer learning technique [6]. Transfer learning method uses a model p
Data Loading...
