Semi-Supervised Domain Adaptation with Latent Diffusion for Pathology Image Classification.

Area of Science:

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

Background:

• Deep learning models in computational pathology struggle with generalization due to domain shift.
• Current domain adaptation methods often fail to use unlabeled data or use image-to-image translation, risking accuracy.
• Domain shift is a significant barrier to deploying AI models in diverse clinical settings.

Purpose of the Study:

• To develop a semi-supervised domain adaptation (SSDA) framework to improve the generalization of computational pathology models.
• To leverage latent diffusion models for generating morphology-preserving, target-aware synthetic pathology images.
• To enhance model performance on unseen target cohorts without compromising performance on the source cohort.

Main Methods:

• A semi-supervised domain adaptation (SSDA) framework was developed using a latent diffusion model.
• The diffusion model was trained on unlabeled data from source and target domains, conditioned on foundation model features, cohort identity, and tissue preparation.
• Synthetic, target-aware images were combined with real, labeled source data to train a downstream classifier for lung adenocarcinoma prognostication.

Main Results:

• The proposed SSDA framework significantly improved performance on a held-out target cohort test set.
• Weighted F1 score increased from 0.611 to 0.706, and macro F1 score improved from 0.641 to 0.716.
• The approach enhanced target-cohort performance without degrading source-cohort performance.

Conclusions:

• Target-aware diffusion-based synthetic data augmentation is a promising approach for improving domain generalization in computational pathology.
• The SSDA framework effectively addresses domain shift by generating realistic and relevant synthetic data.
• This method offers a viable solution for deploying robust AI models across diverse pathology datasets.