• info@icbcongress.com
• English

ساخت اکانت کاربری رایگان

• ورود

INTERNATIONAL CONGRESS OF

CancerGenetics

• Multi-Center Platform for Cancer Detection from Histopathology Images Using Federated Learning

• Aida Yavari Kondori,¹ Hamid Naderi,² Ahmadreza Tavasouli,³ Mona Maftouh,⁴ Masoud Pezeshkirad,⁵ Amir Avan,^6,*
  1. Metabolic Syndrome Research Center, Mashhad University of Medical Sciences, Mashhad, Iran
  2. Metabolic Syndrome Research Center, Mashhad University of Medical Sciences, Mashhad, Iran
  3. Radiology Department, Imam Reza Hospital, Mashhad University Of Medical Sciences, Mashhad, Iran
  4. Radiology Department, Imam Reza Hospital, Mashhad University Of Medical Sciences, Mashhad, Iran
  5. Radiology Department, Imam Reza Hospital, Mashhad University Of Medical Sciences, Mashhad, Iran
  6. Metabolic syndrome Research center, Mashhad University of Medical Sciences, Mashhad, Iran
  
  
  
• Introduction: In connection with the development of deep learning, which became an important sub-field of artificial intelligence, there have been new approaches for analyzing images in medicine, especially in computer-aided diagnosis. The combination of digital pathology and machine learning has created an opportunity for using powerful algorithms that can process a large amount of histopathological images efficiently and consistently. However, the use of AI technologies in pathology in clinical practice is still restricted due to the necessity to store personal information in one place and the issues of confidentiality, data governance, and collaboration between centers. The federated learning approach solves this problem and allows training a shared diagnostic model without sharing raw patients' data across centers. In this study, we aimed to design, build, and deploy a functional federated learning-based multi-center platform for lung and colon cancer detection, integrating deep learning models with a privacy-preserving distributed architecture.
• Methods: Our platform is built using the LC25000 dataset, which comprises 25,000 public color histopathological images divided into five categories: lung adenocarcinoma, lung squamous cell carcinoma, benign lung, colon adenocarcinoma, and benign colon tissue. After performing image pre-processing, we built two independent CNNs models for the classification of lung and colon histopathological images, respectively. In order to facilitate multi-site cooperation, we built a federated learning system that utilizes the Flower framework. Our system architecture consists of one central federated server and two clients in total, where each client holds both lung and colon CNN models in their respective local systems. Using this system architecture, each client is able to build models from their local data without needing to send the actual histopathological images to the central server, but rather updating the server through sending model updates, which get aggregated by the server to build a global model.
• Results: We were able to implement and use a working platform that uses federated learning in order to detect lung and colon cancers using digital histopathology. Two separately trained CNN models are integrated into the distributed multi-client framework, which performs local classification of lung and colon cancers, while simultaneously improving the common global model using the Flower library. The whole framework works as a complete pipeline consisting of local histopathological image pre-processing and classification, followed by secure model aggregation on the server side, without any centralized storage of patient-specific image data. The implemented platform is working and could easily be expanded with new participants and cancer types.
• Conclusion: We present a functional and operational federated learning framework for performing collaborative diagnosis of lung and colon cancers through histopathology images based on deep learning without violating patients' privacy. By using the Flower framework, this framework proves the possibility of developing a multi-center AI-powered pathological diagnosis without the need to pool data in one place. This study provides a practical basis for the future growth of federated digital pathology networks.
• Keywords: Digital Pathology, Federated Learning, Deep Learning, Cancer Detection, Multi-center Platform