Abstract

In recent years, advances in artificial intelligence (AI), particularly deep learning (DL), have transformed medical image analysis. Despite this progress, deploying trained deep-learning models in real-world clinical environments remains challenging due to domain shift problem—the discrepancy between the data distribution used for training and the data encountered during deployment. This issue is especially pronounced in histopathology image classification, where variations in acquisition pipelines across hospitals lead to variation in data distribution of the images. As a result, models trained on a set of source domains struggle to generalize to unseen distributions, resulting in performance degradation. This thesis addresses the challenge of domain shift in histopathology image classification by proposing two schemes that aim to improve domain generalization, each approaching the problem from a different perspective. The first scheme, referred to as PathoWAve, is more concerned with how the classifier model is trained to become robust to unseen domains. It introduces a training strategy where a single classifier model is trained along several parallel trajectories using different augmentations, and the model weights from all trajectories are averaged after each step to stabilize learning procedure. Motivated by the fact that domain shifts
are bounded and continuous in the feature space, the second scheme, referred to as PathoGen, is concerned with enriching the training set with suitable synthetic images. Using a conditional stable diffusion model, PathoGen generates synthetic intermediate-domain images that lie between the original training domains. The resulting expanded training set, comprising both the original and intermediate domains, is then used to train the classifier. By combining these synthetic images with the original data, the training set becomes more continuous across domains in the feature space and accordingly increase the likelihood of the classifier to have seen the images in the target domain. Our results indicate that both PathoWAve and PathoGen lead to significant improvements in generalization capability of the classifier models.