RegionFormer: A Simple Region Transformer for Alzheimer's Disease Diagnosis

149 

Abstract Submission 

Yanling Fu¹, Qi Zhu¹, Wei Shao¹, Wan Peng², Jiashuang Huang³, Daoqiang Zhang¹, Liang Sun¹

¹Nanjing University of Aeronautics and Astronautics, Nanjing, Jiangsu, ²Nanjing University of Aeronautics and Astronautics, Nanjing, China, ³Nantong University, Nantong, Jiangsu

Yanling Fu  
Nanjing University of Aeronautics and Astronautics
Nanjing, Jiangsu

Qi Zhu  
Nanjing University of Aeronautics and Astronautics
Nanjing, Jiangsu

Wei Shao  
Nanjing University of Aeronautics and Astronautics
Nanjing, Jiangsu

Wan Peng  
Nanjing University of Aeronautics and Astronautics
Nanjing, China

Jiashuang Huang  
Nantong University
Nantong, Jiangsu

Daoqiang Zhang  
Nanjing University of Aeronautics and Astronautics
Nanjing, Jiangsu

Liang Sun  
Nanjing University of Aeronautics and Astronautics
Nanjing, Jiangsu

The anatomical structure changes in the regions of the brain are key for understanding brain diseases. However, it is difficult to find slight anatomical structural changes from whole brain images by voxel-level/patch-level deep learning methods. Meanwhile, the recent region-level deep learning methods usually employ multiple sub-networks to learn feature map in each brain region, which are difficult to implement. To address these issues, we propose an effective region-based transformer method for Alzheimer's disease (AD) diagnosis, named RegionFormer. RegionFormer consists of two components:1) a very simple region feature learning network to extract the feature map for each brain region from whole structure magnetic resonance imaging (sMRI), and 2) a transformer-based classifier to capture the dependencies of each brain region for Alzheimer's disease diagnosis.

Region Feature Learning Network. Region feature learning network has two parts，i.e., an image feature learning module and a region-level feature learning module. We first employ the image feature learning module to learn the high-level contextual feature maps of input sMRI image. It's an encoder-decoder architecture, containing six 3D convolutional layers, two max-pooling layers and two 3D deconvolutional layers. Simultaneously, in the region-level feature learning module, the input sMRI image is segmented by a trained segmentation model to obtain 95 brain regions, denoted as L_I. Meanwhile, the output of image feature learning module is fed into a 1×1×1 convolutional layer with 95 channels (i.e., the number of brain regions), denoted as F_BR. Then, we extract region-level feature learning representation based on the label map L_I. Specifically, we first obtain a region-enhanced feature map A_r=(α∗one_hot(L_I)+β)⊗F_BR, where one_hot(•) is one-hot coding and ⊗ is element-wise multiplication. α and β are hyperparameters to adjust the features within and outside the corresponding brain region. Then, the feature map A_r is fed into two group convolutional layers to obtain region-level feature vectors. Finally, the sequence of region-level tokens is used as the input of a transformer-based classifier for Alzheimer's disease diagnosis.
Transformer-based Classifier. In the transformer-based classifier, we stack four transformer encoders to process the region-level features. Similar to ViT, the transformer encoder consists of a multi-head attention (MSA) block and an MLP block. Finally, a softmax function is applied to normalize the outputs.

Our RegionFormer has been evaluated on the baseline sMRI scans of 1193 subjects on the ADNI dataset. The dataset contains 389 AD, 172 pMCI, 232 sMCI, and 400 NC subjects. The accuracy achieved by our RegionFormer in AD vs. NC, pMCI vs. sMCI, pMCI vs. NC, and sMCI vs. NC tasks are 0.983, 0.901, 0.956, and 0.932, which is superior to the state-of-the-art methods. Meanwhile, our RegionFormer finds the AD-related brain regions, which suggests RegionFormer has good pathology interpretability.

The experimental results on ADNI datasets demonstrate that our RegionFormer achieves much better classification performance than several state-of-the-art methods, especially in the relatively challenging task of MCI conversion prediction and early AD diagnosis. Besides, our RegionFormer is easily implemented and extended with the state-of-the-art deep learning framework, segmentation methods, or different brain atlases.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ¹

Executive Function, Cognitive Control and Decision Making

Classification and Predictive Modeling

Anatomical MRI ²

Aging

Amnesia

Cognition

Degenerative Disease

MRI

STRUCTURAL MRI

Thalamus