A new 3D MRI segmentation method based on Generative Adversarial Network and Atrous Convolution. (January 2022)

• Record Type: Journal Article
• Authors: Çelik, Gaffari; Talu, Muhammed Fatih
• Is Part Of: Biomedical signal processing and control. Volume 71(2022)Part A
• Publication Date: 2022-01
• Subjects: Brain segmentation -- 3D MRI scans -- GAN -- ACFP -- PAM; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.103155
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A New 3D MRI Segmentation Method (Vol2SegGAN) is proposed. It uses GAN and ACFP. It provided the highest segmentation accuracy for CSF and GM segmentation. It provided the second highest for WM segmentation. It provided the best segmentation accuracy according to VS and HD metrics. It provided the second best according to Dice in eight tissues segmentation. It has the best segmentation time in all experiments made. Abstract: Background and aim: Although many algorithms have been proposed to segment brain structures in MRI scans, comparison of different algorithms in the same data set is rarely done. Many methods still run on privately held data and include comparisons with previous versions. The purpose of this study is to introduce a new Generative Adversarial Network (GAN) based segmentation architecture. Brain tissues on 3D-MRI scans with T1w modality were segmented into three (GM, WM, CSF) and eight (CGM, BG, WM, WMH, CF, VE, CE ve BS) parts. Methods: The proposed approach (Vol2SegGAN) consists of two parts: preprocessing and segmentation. The preprocessing includes extraction of the brain region, editing of labels in the datasets, MNI152 registration, clipping/sampling. The segmentation part is carried out by the collaboration of a generator (includes ACFP and PAM modules) and a discriminator (distinguishes real/fake) architectures. Conclusions: In the three part segmentation process, the proposed method showed the best segmentation success in CSF (Dice = 0.739, VS = 0.967), GM (Dice = 0.878, HD = 2.378) and WM (HD = 2.105) tissues, and the second best segmentation success in WM (Dice = 0.793, VS = 0.972) tissue according to Dice and VS metrics. Similarly, in the segmentation of eight parts, it is seen that it has the best success according to VS and HD metrics and the second best according to Dice metric. Vol2SegGAN, which has fewer parameters (6, 883 mil.) than existing architectures, has an average of 11–12 s (CPU) or 0–1 s (GPU) to segment a sample 3D-MRI. Implementation codes of the proposed architecture are available on the github page 1 .