TDABNet: Three-directional attention block network for the determination of IDH status in low- and high-grade gliomas from MRI. (May 2022)

• Record Type: Journal Article
• Authors: Ai, Lingmei; Bai, Wenhao; Li, Mengge
• Is Part Of: Biomedical signal processing and control. Volume 75(2022)
• Publication Date: 2022-05
• Subjects: Isocitrate dehydrogenase -- Deep learning -- Three-dimensional convolutional neural network -- Attention mechanism -- Three-directional attention block network (TDABNet); 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.2022.103574
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: A novel Three-Directional Attention Block (TDAB) is introduced. Different from traditional attention blocks, our proposed TDAB can capture the long-range dependencies that are helpful for predicting IDH mutation status. We propose a 3D convolutional network architecture based on multi-scale feature fusion algorithm. Different from other studies view 3D MRI as 2D slice images, we feed 3D MR images into the TDABNet. We propose a Z-Directional Attention Block (ZAB) and Feature Separation Block (FSB) based on the 3D multi-scale convolutional network architecture. Abstract: The isocitrate dehydrogenase (IDH) mutation in low- and high-grade gliomas have proven to be the critical molecular biomarker associated with better prognosis. Although the determination of the IDH status of these neoplasms prior to surgical intervention is considered beneficial for prognosis, this information is currently only available after surgical removal of the tissue. At present, most studies have proved the efficiency of deep learning technology in noninvasive diagnosing IDH status. However, there are still some shortages. Firstly, they only input the 2D slices of gliomas into the network, ignoring the significant amount of extra information of gliomas in the third dimension. Secondly, because glioma is a heterogeneous three-dimensional volume with complex imaging features, it is still a challenge for traditional CNN to learn the features that help predict IDH status from magnetic resonance imaging (MRI). To address these issues, we propose a Three-Directional Attention Block Network (TDABNet) based on a three-dimensional convolutional neural network (3D CNN), which can accurately determine the IDH status in gliomas from 3D MRI. The performance of TDABNet was validated in a dataset of 235 patients with low- and high-grade gliomas and the area under the operating characteristic curve (AUC) of IDH status prediction is 96. 44%. It is proved by experiment that TDABNet can accurately predict the IDH status of gliomas.