A 3D efficient and essentialized swin transformer network for alzheimer’s disease diagnosis

Abstract

Deep learning methods (e.g., convolutional neural networks, CNNs) have been widely applied to Alzheimer’s disease diagnosis based on structural magnetic resonance imaging (sMRI) data. However, CNN-based methods face significant Limitations in capturing the global feature distribution of the whole brain. Transformer-based models have shown promise in addressing this issue, but they often sacrifice local feature sensitivity. Moreover, the large number of parameters in Transformer-based models results in a strong dependence on large-scale datasets, which is difficult to satisfy in real-world 3D medical imaging scenarios. Through comprehensive consideration, we propose a 3D Efficient and Essentialized Swin Transformer Network (E2STN) to strike a balance between being lightweight and comprehensive feature extraction, thereby boosting Alzheimer’s disease diagnosis performance in 3D dataset scenarios. Specifically, E2STN includes four modules: an Efficient Swin Transformer (EST) module for identifying global structural information and being lightweight to reduce reliance on large-scale datasets, which is a novel task-oriented Transformer architecture; a Focused Feature Enhancement Convolution Unit (FFE-CU) for enhancing lesion details, thereby compensating for the limited perception of fine-grained pathological information by the Transformer; a Disease Risk Map generator (DRMg) for visualizing pathological regions; and an ROI-based classifier for precise categorization. Our proposed method has been validated by two diagnosis tasks (i.e., Alzheimer’s disease diagnosis and mild cognitive impairment conversion prediction) on the ADNI dataset. Compared to several state-of-the-art methods, our model demonstrates superior performance.