Scholarlink

<p>Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by the loss of dopamine-producing neurons, resulting in motor impairments such as tremors, muscle rigidity, slowed movements, and changes in handwriting. Although there is currently no cure for PD, early diagnosis can facilitate timely intervention and improve symptom management. Handwriting analysis has emerged as a promising non-invasive approach for identifying motor abnormalities associated with the disease. This study presents DrawNet, an optimizer-aware transfer learning framework for automated Parkinson’s disease detection using hand-drawn spiral and wave patterns. Experiments were conducted on a publicly available handwriting dataset containing samples from both individuals with Parkinson’s disease and healthy controls. To address the limited dataset size and enhance model generalization, a structured preprocessing pipeline incorporating normalization and offline data augmentation techniques, including rotation, shifting, zooming, and shearing, was employed. Nine state-of-the-art pre-trained convolutional neural networks–ResNet50, ResNet101, DenseNet201, VGG19, InceptionV3, InceptionResNetV2, MobileNetV2, Xception, and NASNetMobile were evaluated for binary classification. Three optimization algorithms, namely SGDM, Adam, and RMSprop, were systematically investigated to assess their influence on learning performance across different architectures. Model performance was assessed using accuracy, precision, recall, and F1-score under multiple train–validation–test split configurations, complemented by 5-fold cross-validation and statistical significance analysis. Among all evaluated combinations, MobileNetV2 with RMSprop achieved the best overall performance, attaining an average test accuracy of 97.24% under the 80:10:10 data split while maintaining consistent results across repeated experiments and cross-validation folds. These findings highlight the potential of transfer learning-based handwriting analysis as a cost-effective and non-invasive decision-support tool for early Parkinson’s disease screening. Nevertheless, further validation on larger and more diverse clinical datasets is required to establish broader real-world applicability.</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

Drawnet an optimizer aware transfer learning framework for Parkinson’s disease detection from hand drawn spirals and waves

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月16日

查看全文
  • Makkapati Lakshmi Prasanna,
  • Thulasi Bikku,
  • Tony Jan,
  • Chandan Kumar

摘要

Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by the loss of dopamine-producing neurons, resulting in motor impairments such as tremors, muscle rigidity, slowed movements, and changes in handwriting. Although there is currently no cure for PD, early diagnosis can facilitate timely intervention and improve symptom management. Handwriting analysis has emerged as a promising non-invasive approach for identifying motor abnormalities associated with the disease. This study presents DrawNet, an optimizer-aware transfer learning framework for automated Parkinson’s disease detection using hand-drawn spiral and wave patterns. Experiments were conducted on a publicly available handwriting dataset containing samples from both individuals with Parkinson’s disease and healthy controls. To address the limited dataset size and enhance model generalization, a structured preprocessing pipeline incorporating normalization and offline data augmentation techniques, including rotation, shifting, zooming, and shearing, was employed. Nine state-of-the-art pre-trained convolutional neural networks–ResNet50, ResNet101, DenseNet201, VGG19, InceptionV3, InceptionResNetV2, MobileNetV2, Xception, and NASNetMobile were evaluated for binary classification. Three optimization algorithms, namely SGDM, Adam, and RMSprop, were systematically investigated to assess their influence on learning performance across different architectures. Model performance was assessed using accuracy, precision, recall, and F1-score under multiple train–validation–test split configurations, complemented by 5-fold cross-validation and statistical significance analysis. Among all evaluated combinations, MobileNetV2 with RMSprop achieved the best overall performance, attaining an average test accuracy of 97.24% under the 80:10:10 data split while maintaining consistent results across repeated experiments and cross-validation folds. These findings highlight the potential of transfer learning-based handwriting analysis as a cost-effective and non-invasive decision-support tool for early Parkinson’s disease screening. Nevertheless, further validation on larger and more diverse clinical datasets is required to establish broader real-world applicability.