Automatic segmentation of the brain stroke lesions from MR flair scans using improved U-net framework. (September 2022)

• Record Type: Journal Article
• Authors: Khezrpour, Samrand; Seyedarabi, Hadi; Razavi, Seyed Naser; Farhoudi, Mehdi
• Is Part Of: Biomedical signal processing and control. Volume 78(2022)
• Publication Date: 2022-09
• Subjects: Medical image segmentation -- Deep learning -- Convolutional networks -- And stroke lesion images; 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.2022.103978
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: We present a methodology for autonomously segmenting stroke lesions in the FLAIR modality images. A deep supervised U-Net architecture is used in our proposed network, which incorporates Blocks made up of five parallel layers. We assessed the proposed framework On the MICCAI 2015 Ischemic Stroke Lesion Segmentation dataset (ISLES2015) Challenge. Experiments show that compared to traditional machine learning methods, proposed method shows better performance. We anticipate this methodology has a lot of potential in the medical profession for diagnosing and treating other brain disorders as well as organ segmentation. Abstract: Background: Magnetic resonance imaging (MRI) can reliably diagnose ischemic stroke. Stroke is an acute vascular illness of the brain that can lead to long-term death and disability. The majority of stroke patients have acute ischemic lesions. These stroke lesions are treatable underneath correct diagnosing and treatment. Despite the fact that fluid attenuation inversion recovery (FLAIR) images are susceptible to detecting these types of lesions, clinicians find it difficult to locate and measure them manually. New method: In this research, we present a methodology for autonomously segmenting stroke lesions in the FLAIR modality images. A deep supervised U-Net architecture is used in our proposed network, which incorporates Blocks made up of five parallel layers. Results & comparison with existing methods: We assessed the proposed framework On the MICCAI 2015 Ischemic Stroke Lesion Segmentation dataset (ISLES2015) Challenge. In conclusion, the dice coefficient attained a mean accuracy of 0.89. Conclusions: Experiments show that compared to traditional machine learning methods, proposed method shows better performance. The experiment results have already confirmed that the proposed U-Net model is a better tool for dealing with segmentation problems that are related to others on similar datasets.