Publication Type : Book Chapter
Publisher : CRC Press
Source : Deep Learning for Biomedical Applications, CRC Press, p.83–100 (2021)
Url : https://www.taylorfrancis.com/chapters/edit/10.1201/9780367855611-5/performance-analysis-deep-learning-models-biomedical-image-segmentation-saj-sachin-sowmya-soman
ISBN : 9780367855611
Campus : Coimbatore
School : School of Engineering
Center : Center for Computational Engineering and Networking
Department : Electrical and Electronics
Year : 2021
Abstract : Image segmentation is the process of partitioning an image into multiple meaningful segments, which can be easily analyzed. Image segmentation is one of the most common tasks in computer vision. Cancer is termed as abnormal growth of cells and it can be either malignant or benign. Since the cancer rates in the world are increasing at an alarming rate and if it is found at earlier stages, chances of getting the cancer cured are very high. Since manual annotation of these cancer regions by doctors from medical imaging like X-Ray, MRI, etc., is highly time-consuming and can be prone to human errors, automatic and accurate tumor segmentation is of great interest. In recent years, deep learning is showing great progress in most of the applications and thus motivated us to use in the biomedical image segmentation task. The data that we have considered in this chapter are brain tumor (preprocessed BraTs 2015) and skin tumor (International Skin Imaging Collaboration (ISIC) 2018), as it is dangerous and as well as the rate of people getting affected is also increasing around the world. For our biomedical image segmentation, we proposed to use three architectures: (a) SegAN, (b) SegNet, and (c) U-Net. Segmentor adversarial network (SegAN) belongs to generative adversarial networks (GAN) family and other two architectures are commonly used for image segmentation task. Initially, performance analysis of SegAN was done using BraTs 2017 and ISIC 2018 data set. Based on the standard evaluation parameters such as Jaccard index and Dice score, the segmented/predicted output is evaluated. Then to increase the segmentation accuracy, U-Net and SegNet are used for skin and brain tumor image segmentation. In terms of Jaccard index, our proposed work achieved comparable performances, and in terms of Dice score, our proposed work outperformed the existing approaches in skin lesion image segmentation by achieving a maximum Dice score of 87.56% with U-Net architecture. For brain tumor segmentation, our proposed work was able to outperform the existing approaches by achieving a maximum Dice score of 85.01% with U-Net architecture. Lastly, segmentation output by all three architectures was compared for skin lesion segmentation and found that U-Net gave the best segmentation output followed by SegAN. In case of SegNet, the segmentation output is smooth-edged. In terms of learnable parameters, SegAN has the highest and U-Net has the least, when compared with all the three architectures. Thus, the proposed chapter concludes that U-Net achieved the best parameter results for both skin and brain tumor image segmentation task when compared with SegAN and SegNet.
Cite this Research Publication : T. K. Saj Sachin, Sowmya V., and Dr. Soman K. P., “Performance Analysis of Deep Learning Models for Biomedical Image Segmentation”, in Deep Learning for Biomedical Applications, 1st Editionst ed., CRC Press, 2021, pp. 83–100.