MRI-based model for MCI conversion using deep zero-shot transfer learning

This study describes a deep zero-shot transfer learning model (DZTLM) for predicting mild cognitive impairment (MCI) in patients with Alzheimer's disease (AD). The proposed DZTLM combines ResNet and deep subdomain adaptation network (DsAN) blocks with a simple data augmentation and transfer technique, Elastic-Mixup. We test the DZTLM using 3D gray matter images segregated from structural MRI as input. Ablation experiments are conducted to evaluate the proposed model and compare it with existing approaches. Experiments demonstrate that the DsAN network coordinating Elastic-Mixup enhances the accuracy of MCI-AD prediction by more than 18% compared with a standard 3D ResNet50 classifier. The Elastic-Mixup technique contributes more than 16% to this increase in prediction accuracy. Elastic-Mixup also enhances the sensitivity of recognition for stable MCI. When labeled samples are scarce, the unsupervised DZTLM outperforms a semi-supervised transfer learning model. The DZTLM achieves comparable outcomes to existing models despite the absence of tagged MRI data.

Automated summary

This study presents a deep zero-shot transfer learning model for predicting mild cognitive impairment in Alzheimer's disease patients, which achieves improved accuracy compared to existing approaches.