Scholarlink

<p>ParaFlow presents a semi-layered hybrid neural architecture integrating Liquid Time-Constant Networks (LTCs) and Neural Spline Flows (NSFs) for the automated diagnosis of pemphigus vulgaris (PV) from histopathology images. The LTC branch captures continuous tissue dynamics, including fine-scale lesion evolution, and the NSF branch normalizes feature embeddings via invertible transforms. On a benchmark set of 500 labeled histopathology slides, ParaFlow reported 95.2% accuracy, 94.8% precision, 95.0% recall, and 94.9% F1-score, beating eight baseline models. For instance, ResNet50 scored 89.1%, DenseNet121 scored 90.3%, and Capsule Networks scored 91.5%. ParaFlow also cut inference latency by 1.8× over Transformer-based models, while showing both high accuracy and low latency. These findings confirm that ParaFlow correctly extracts local and global tissue features, enhancing discrimination among fine pemphigus lesion patterns. The method presents a strong and stable tool for clinical diagnosis, resolving the deficiency of current deep learning methods in precisely identifying pemphigus subtypes.</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

ParaFlow: a semi-layered Liquid Time-Constant (LTC) and Neural Spline Flow (NSF) network for automated diagnosis of pemphigus vulgaris

  • 期刊论文

  • 发表日期：

    2025年12月19日

查看全文
  • Sonam Dubey,
  • Kamlesh Kumar Singh,
  • Prashant Kumar Mishra

摘要

ParaFlow presents a semi-layered hybrid neural architecture integrating Liquid Time-Constant Networks (LTCs) and Neural Spline Flows (NSFs) for the automated diagnosis of pemphigus vulgaris (PV) from histopathology images. The LTC branch captures continuous tissue dynamics, including fine-scale lesion evolution, and the NSF branch normalizes feature embeddings via invertible transforms. On a benchmark set of 500 labeled histopathology slides, ParaFlow reported 95.2% accuracy, 94.8% precision, 95.0% recall, and 94.9% F1-score, beating eight baseline models. For instance, ResNet50 scored 89.1%, DenseNet121 scored 90.3%, and Capsule Networks scored 91.5%. ParaFlow also cut inference latency by 1.8× over Transformer-based models, while showing both high accuracy and low latency. These findings confirm that ParaFlow correctly extracts local and global tissue features, enhancing discrimination among fine pemphigus lesion patterns. The method presents a strong and stable tool for clinical diagnosis, resolving the deficiency of current deep learning methods in precisely identifying pemphigus subtypes.