Pseudo-Data Based Self-Supervised Federated Learning for Classification of Histopathological Images

Area of Science:

• Medical image analysis
• Computational pathology
• Artificial intelligence in healthcare

Background:

• Computer-aided diagnosis (CAD) models for cancer improve accuracy but struggle with generalization due to data inconsistencies across institutions.
• Federated learning (FL) offers a solution by training models across multiple centers without sharing raw data.

Purpose of the Study:

• To propose a pseudo-data based self-supervised federated learning (FL) framework (SSL-FL-BT) to enhance the diagnostic accuracy and generalization of CAD models for cancer.
• To address the generalization challenges of CAD models trained on single-center histopathological data.

Main Methods:

• Generation of pseudo histopathological images at each center, retaining center-specific properties without privacy risks.
• Utilizing pseudo images for self-supervised learning (SSL) on a central server to pre-train the global model backbone.
• Implementing a multi-task SSL to learn both center-specific and common representations.
• Employing a novel Barlow Twins based FL (FL-BT) for contrastive learning to improve local training and global model optimization.

Main Results:

• The proposed SSL-FL-BT framework demonstrated significant improvements in both diagnostic accuracy and generalization capabilities.
• Experimental validation on four public histopathological image datasets confirmed the framework's effectiveness.

Conclusions:

• The SSL-FL-BT framework effectively enhances CAD models for cancer diagnosis by leveraging pseudo-data, SSL, and FL.
• This approach successfully mitigates the generalization problem in multi-center histopathological image analysis.