A Swin Transformer Approach for Amyloid PET Image Analysis in Neurodegenerative Disease Detection

Positron emission tomography (PET) imaging techniques are fundamental in the detection of neurodegenerative diseases, particularly Alzheimer’s disease, by detecting radiopharmaceutical absorption in brain tissues. While convolutional neural networks (CNNs) have shown promise in automated detection of amyloid plaques, the hierarchical architecture of vision transformers offers potential advantages for medical image analysis. This research proposes a novel approach using Swin Transformer architecture to analyze a dataset of PET images from patients undergoing diagnosis for neurodegenerative pathologies. Thanks to the use of the shifted window attention mechanism, the model is able to effectively capture both local characteristics and global relationships present in brain images. After GPU-accelerated training, the Swin Transformer model demonstrated enhanced capability in recognizing amyloid plaque accumulation compared to traditional CNN approaches. Based on prior research in transformer-based medical image analysis, we hypothesize improved accuracy (89–92%) and recall (80–84%) metrics for amyloid detection, addressing the challenges of inter-reader variability in clinical settings. The proposed model offers significant potential for computer-aided diagnosis systems that can assist specialists in early and accurate detection of neurodegenerative conditions.