Improvements in dementia classification for brain SPECT volumes using vision transformer and the Brodmann areas

This study proposes a vision transformer (ViT)-based model for dementia classification, able to classify representative dementia with Alzheimer's disease, dementia with Lewy bodies, frontotemporal dementia, and healthy controls using brain single-photon emission computed tomography (SPECT) images. The proposed method allows for an input based on the anatomical structure of the brain and the efficient use of five different SPECT images.

The proposed model comprises a linear projection of input patches, eight transformer encoder layers, and a multilayered perceptron for classification with the following features: 1. diverse feature extraction with a multi-head structure for five different SPECT images; 2. Brodmann area-based input patch reflecting the anatomical structure of the brain; 3. cross-attention to fusion of diverse features.

The proposed method achieved a classification accuracy of 85.89% for 418 SPECT images from real clinical cases, significantly outperforming previous studies. Ablation studies were conducted to investigate the validity of each contribution, in which the consistency between the model's attention map and the physician's attention region was analyzed in detail.

The proposed ViT-based model demonstrated superior dementia classification accuracy compared to previous methods, and is thus expected to contribute to early diagnosis and treatment of dementia using SPECT imaging. In the future, we aim to further improve the accuracy through the incorporation of patient clinical information.