Spatial-Frequency dual-branch attention model for determining KRAS mutation status in colorectal cancer with T2-weighted MRI. (September 2021)

• Record Type: Journal Article
• Authors: Ma, Yulan; Wang, Jiawen; Song, Kai; Qiang, Yan; Jiao, Xiong; Zhao, Juanjuan
• Is Part Of: Computer methods and programs in biomedicine. Volume 209(2021)
• Publication Date: 2021-09
• Subjects: Colorectal cancer -- KRASmutation status -- Spatial-domain -- Frequency-domain -- Attention module -- T2-weighted MRI; Medicine -- Computer programs -- Periodicals; Biology -- Computer programs -- Periodicals; Computers -- Periodicals; Medicine -- Periodicals; Médecine -- Logiciels -- Périodiques; Biologie -- Logiciels -- Périodiques; Biology -- Computer programs; Medicine -- Computer programs; Periodicals; Electronic journals; 610.28
• DOI: 10.1016/j.cmpb.2021.106311
• Languages: English
• ISSNs: 0169-2607

Abstract

Highlights: We designed a deep CNN model that combines spatial and frequency domains information to determine KRAS mutation status of CRC patients. Through the way of channel rearrangement, the frequency-domain information of the same frequency are rearranged into pseudo 3D-DCT cube columns without losing any information. A channel enhancement attention module (CEAM) is proposed to make the model more purposeful for training. Transfer learning is used in the spatial-domain branch to alleviate the problem of a small-scale dataset. Abstract: Background and Objective: Identifying the KRAS mutation status accurately in medical images is very important for the diagnosis and treatment of colorectal cancer. Despite the substantial progress achieved by existing methods, it remains challenging due to limited annotated dataset, large intra-class variances, and a high degree of inter-class similarities. Methods: To tackle these challenges, we propose a spatial-frequency dual-branch attention model (SF-DBAM) to determine the KRAS mutation status of colorectal cancer patients using a limited T2-weighted MRI dataset. The dataset contains 169 wild-type patients (2151 images) and 137 mutation-type patients (1666 images). The first branch utilizes part of the pre-trained Xception model to capture spatial-domain information and alleviate the small-scale dataset problem. The second branch builds frequency-domain information into cube columns using block-based discrete cosine transform and channel rearrangement. Then the cube columns are fed into convolutional long short-term memory (convLSTM) to explore the effective information between the reconstructed frequency-domain channels. Also, we design a channel enhanced attention module (CEAM) at the end of each branch to make them focus on the lesion areas. Finally, we concatenate the two branches and output the classified results through fully connected layers. Results: The proposed method achieves 88.03% overall accuracy with AUC of 94.27% and specificity of 90.75% in 10-fold cross-validation, which is better than the current non-invasive methods for determining KRAS mutation status. Conclusions: We believe that the proposed method can assist physicians to diagnose the KRAS mutation status in patients with colorectal cancer, and other medical problems can benefit from the spatial and frequency domains information.