Transfer Learning from Vision Transformer to Parkinson's Disease Offline Handwriting Detection
Huaixin Liang, Tao Zhang, Lin Li, Bo Liu, Bin Liu, Chen Shi, BaoHong Mi · 2024
Handwriting has emerged as a crucial method for detecting Parkinson's disease (PD). The use of offline handwriting images eliminates the need for special equipment, making it more adaptable for PD detection. This paper presents a Transfer Learning from Vision Transformer (TLViT) approach for offline handwriting PD detection. Transfer learning addresses the issue of limited sample size in Parkinson's offline handwriting images. Additionally, the multi-head attention module in Vision Transformer allows the network to focus more accurately on changes in handwriting within offline images. To enhance overall model prediction performance, this paper combines these advantages to successfully complete the task of PD detection. Experiments conducted on three handwriting datasets validate the effectiveness of the proposed method. The TLViT method achieves 97.31% accuracy for Spiral tasks in the NewHandPD dataset and an average accuracy of 96.96% across all tasks within that dataset. These results demonstrate that the proposed method improves prediction performance and efficiency compared to existing methods. In conclusion, the TLViT method offers a novel and powerful tool for detecting PD through offline handwriting images.”