Multi-Modal Heterogeneous Encoder-Decoder for Alzheimer's Disease Anomaly Detection

Alzheimer's disease (AD) stands as the primary cause of dementia, and early diagnosis is crucial for treatment and prevention. This article proposes a novel unsupervised anomaly detection method that integrates structural magnetic resonance imaging (sMRI) with positron emission tomography (PET) data to construct multi-modal brain imaging datasets. This method employs a heterogeneous encoder-decoder framework, using pre-trained ResNet50 as the encoder and ResNeXt50 as the decoder, with the goal of enhancing the detection capabilities for brain abnormalities. Furthermore, the encoder incorporates a hybrid attention mechanism and a multi-scale feature fusion module to boost the capacity to extract and represent complex medical image features. Experimental results using data from the ADNI database demonstrate that our model surpasses current technologies in AUC, accuracy, sensitivity, and specificity, thereby confirming the efficacy of multi-modal data fusion and the heterogeneity of the encoder-decoder framework in the early classification and detection of AD. Additionally, the study relies solely on readily available normal (healthy) images for training, circumventing the high costs associated with the collection and labeling of abnormal (unhealthy) images. This research offers novel insights and methodologies for the early classification and diagnosis of Alzheimer's disease.