DualCoAg-Net: A Federated Contrastive-Bayesian Multi-Task Framework for Robust and Explainable Skin Lesion Segmentation and Classification

DualCoAg-Net: A Federated Contrastive-Bayesian Multi-Task Framework for Robust and Explainable Skin Lesion Segmentation and Classification

Training medical AI models traditionally requires centralizing sensitive patient images across hospitals — a major privacy, legal, and logistical barrier. This project, DualCoAg-Net, resolves this with a state-of-the-art federated learning framework for skin lesion analysis that trains a shared model across distributed clinical sites without any patient data leaving its source institution. The novel architecture integrates three advanced techniques: contrastive learning to learn robust lesion representations from limited labeled data, Bayesian modeling to quantify prediction uncertainty — critical for clinical trust — and multi-task learning that simultaneously performs lesion segmentation and malignancy classification, outperforming single-task baselines. A dual-aggregation strategy intelligently merges client updates to handle non-IID data distributions across hospitals. Every prediction includes confidence intervals, attention-based explainability maps, and uncertainty warnings for cases needing specialist review. The system is validated on the ISIC skin lesion benchmark with simulated multi-client federated setups, reporting Dice, IoU, and AUC metrics alongside communication-efficiency analysis. Implemented with PyTorch and the Flower federated learning framework, this is an elite federated learning healthcare project for postgraduate research and privacy-preserving machine learning teams.

Components

Python 3.8+
PyTorch
Flower federated learning framework
Contrastive learning modules
Bayesian uncertainty layers
ISIC benchmark datasets

Key Features

• Federated training across simulated hospitals without moving patient data
• Contrastive learning for robust lesion representations from limited labels
• Bayesian uncertainty quantification for clinical trust
• Joint segmentation and malignancy classification outperforming single-task baselines
• Dual aggregation strategy handling non-IID data across clients

Applications

Postgraduate research, hospitals building collaborative AI, and ML engineers exploring privacy-preserving machine learning.