Scholarlink

<p>Parkinson’s disease (PD) is a prevalent neurodegenerative disorder in which early diagnosis plays a crucial role in slowing disease progression. Handwriting analysis has recently emerged as a promising, non-invasive, and low-cost screening tool for detecting tremor-induced micro-changes in handwriting. However, the limited size and diversity of PD datasets often lead to poor generalization of deep learning models, restricting their clinical usability. To address these challenges, we propose a lightweight dual-branch diagnostic framework–Bidirectional Dynamic Interaction Network (BDINet). BDINet integrates handcrafted SIFT keypoint features with deep attention-based CoorLGNet features, enabling effective bidirectional interaction through the Multi-angle Feature Interaction Module (MFIM) and refined fusion via the Interactive Fusion Module (IFF). This design enhances the model’s sensitivity to subtle handwriting variations, such as stroke inflection points, curvature, and angular fluctuations, which are indicative of motor instability in PD patients. Experiments on the HandPD, NewHandPD, and Short Text datasets demonstrate accuracies of 99.18%, 97.22%, and 97.73%, respectively–significantly outperforming state-of-the-art models. Despite its high precision, BDINet remains computationally efficient (10.12 GFLOPs) and easy to deploy on standard PCs or edge devices. These results highlight BDINet’s strong clinical applicability and practical feasibility as an accurate, lightweight, and interpretable solution for early Parkinson’s disease screening. Our code is available at <a href="https://github.com/BM-AI-Lab/BDINet">https://github.com/BM-AI-Lab/BDINet</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Fusion of bidirectional dynamic interaction key features for micro-change recognition in Parkinsonian handwriting

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Xiaobei Yu,
  • Xuesen Niu,
  • Haiyang Meng,
  • Xinglin Yu,
  • Aite Zhao

摘要

Parkinson’s disease (PD) is a prevalent neurodegenerative disorder in which early diagnosis plays a crucial role in slowing disease progression. Handwriting analysis has recently emerged as a promising, non-invasive, and low-cost screening tool for detecting tremor-induced micro-changes in handwriting. However, the limited size and diversity of PD datasets often lead to poor generalization of deep learning models, restricting their clinical usability. To address these challenges, we propose a lightweight dual-branch diagnostic framework–Bidirectional Dynamic Interaction Network (BDINet). BDINet integrates handcrafted SIFT keypoint features with deep attention-based CoorLGNet features, enabling effective bidirectional interaction through the Multi-angle Feature Interaction Module (MFIM) and refined fusion via the Interactive Fusion Module (IFF). This design enhances the model’s sensitivity to subtle handwriting variations, such as stroke inflection points, curvature, and angular fluctuations, which are indicative of motor instability in PD patients. Experiments on the HandPD, NewHandPD, and Short Text datasets demonstrate accuracies of 99.18%, 97.22%, and 97.73%, respectively–significantly outperforming state-of-the-art models. Despite its high precision, BDINet remains computationally efficient (10.12 GFLOPs) and easy to deploy on standard PCs or edge devices. These results highlight BDINet’s strong clinical applicability and practical feasibility as an accurate, lightweight, and interpretable solution for early Parkinson’s disease screening. Our code is available at https://github.com/BM-AI-Lab/BDINet.