Scholarlink

<p>The rapid growth of medical imaging data has provided great potential to automate image classification, especially in prostate cancer diagnosis. Digital pathology and whole-slide imaging (WSI) produce high-resolution data that is a prerequisite for precise Gleason grading. Yet, this progress brings its own set of challenges, centered on privacy and data sensitivity, institutional data heterogeneity, and the sustainability of computational resources. In this study, ProstAI-Net, a novel federated deep learning system, is introduced to address these challenges and maintain high diagnostic performance while ensuring data privacy in multi-institutional networks. Employing a combination of convolutional neural networks and federated learning methods such as federated averaging and gradient noise injection, ProstAI-Net maintains patient confidentiality since users’ medical data never leaves their local institutions. As a result, it reduces those concerns of a central storage as well as a shared area for data, leading to increased security. Furthermore, by exploiting decentralized training, the system relieves the computation burden of centralized servers, which in turn reduces 48% in Carbon emissions with respect to data processing. Moreover, ProstAI-Net also uses multi-modal feature fusion: the morphological patterns, GLCM, and optical density measures are combined with WSI features to further improve the classification performance. The framework comes with the attention-based visualization module that guarantees explainability and promotes clinician trust to reveal image regions important for the model predictions. Experiments on the multi-institutional distributed datasets of three hospitals show that ProstAI-Net achieves an average AUC of 0.96, and is an accurate, interpretable, and privacy-preserving approach for prostate cancer image classification. The results demonstrate that ProstAI-Net is a secure, interpretable, and sustainable model for medical image classification and can offer scalable AI solutions with smooth integration into the clinical practice following privacy &amp; security protocols. Our work provides a stepping stone for the development of privacy-preserving AI systems in healthcare and paves the way for future research on adaptive and secure health information systems.</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

ProstAI-Net: A Federated Privacy-Preserving Deep Learning Framework for Sustainable Prostate Cancer Image Classification in Digital Pathology

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月9日

查看全文
  • Muhammad Arshad,
  • Chengliang Wang,
  • Refka Ghodhbani,
  • Jamshed Ali Shaikh,
  • Nihal Abuzinadah,
  • Khaled Alnowaiser,
  • Yunyoung Nam,
  • Yongwon Cho

摘要

The rapid growth of medical imaging data has provided great potential to automate image classification, especially in prostate cancer diagnosis. Digital pathology and whole-slide imaging (WSI) produce high-resolution data that is a prerequisite for precise Gleason grading. Yet, this progress brings its own set of challenges, centered on privacy and data sensitivity, institutional data heterogeneity, and the sustainability of computational resources. In this study, ProstAI-Net, a novel federated deep learning system, is introduced to address these challenges and maintain high diagnostic performance while ensuring data privacy in multi-institutional networks. Employing a combination of convolutional neural networks and federated learning methods such as federated averaging and gradient noise injection, ProstAI-Net maintains patient confidentiality since users’ medical data never leaves their local institutions. As a result, it reduces those concerns of a central storage as well as a shared area for data, leading to increased security. Furthermore, by exploiting decentralized training, the system relieves the computation burden of centralized servers, which in turn reduces 48% in Carbon emissions with respect to data processing. Moreover, ProstAI-Net also uses multi-modal feature fusion: the morphological patterns, GLCM, and optical density measures are combined with WSI features to further improve the classification performance. The framework comes with the attention-based visualization module that guarantees explainability and promotes clinician trust to reveal image regions important for the model predictions. Experiments on the multi-institutional distributed datasets of three hospitals show that ProstAI-Net achieves an average AUC of 0.96, and is an accurate, interpretable, and privacy-preserving approach for prostate cancer image classification. The results demonstrate that ProstAI-Net is a secure, interpretable, and sustainable model for medical image classification and can offer scalable AI solutions with smooth integration into the clinical practice following privacy & security protocols. Our work provides a stepping stone for the development of privacy-preserving AI systems in healthcare and paves the way for future research on adaptive and secure health information systems.