Image Feature Extraction with Fourier Transform and Multi-task Swin-Transformer for Alzheimer’s Disease Prediction and Detection

Alzheimer’s disease (AD) patients often miss the optimal treatment time due to their late onset and unclear early clinical manifestations. There is little difference between the CT images of the brain of AD patients and normal people, and the efficiency of manual diagnosis is low. However, classic models (such as Resnet, Swin Transformer, CNN, etc.) do not have strong targeting for brain disease diagnosis, and they are more inclined towards general classification tasks, so there is great room for improving the accuracy of the model. This article (FTM Swin) is based on a learnable mask mechanism with algorithm combined with SVM and swin-transformer, to filter and extract features in the frequency domain graph of 3D images. Fast Fourier transform (FFT) is used to achieve the interaction and combination of frequency domain and spatial domain feature matrices, ultimately completing the classification task of data points in the time domain, which simplifies the feature extraction process and makes the classification boundary of the classifier more accurate. Subsequently, this article utilized the ADNI brain imaging dataset and trained the previous classifier model using a dataset annotated with the hippocampus region, which plays a crucial role in Alzheimer’s disease lesions. This further enhanced the model’s ability to detect lesion sites and provided information on the stage of the condition. Compared with the other eight advanced models for image classification tasks in the past three years, FTM-Swin has an accuracy of 97.18% on the validation set, with the best performance.