How to Make AI Models in Digital Pathology Perform More Reliably Across Different Sources

What is it about?

This publication is a survey about a critical issue in AI for medical imaging, especially in digital pathology (the study of disease using digital scans of tissue). It focuses on a challenge called domain shift, which happens when AI models trained in one hospital or lab don’t work well in another due to differences in image scanners and staining methods. That's only one way that domain shift can happen, known as covariate shift, where the appearance of the input data changes between domains. For example, the same tissue scanned by two different scanners might look noticeably different due to variations in resolution or color profiles. But domain shift isn’t just about how images look. There are three other important types: - Prior shift occurs when the distribution of labels changes between domains. For instance, if your training data has a balanced number of cancerous and non-cancerous slides, but your deployment data has mostly non-cancerous ones, your model might become biased or inaccurate. - Posterior shift happens when the relationship between features and labels changes, often due to subjective annotations. A common example is mitosis detection, where two expert pathologists might label the same image differently, leading to inconsistent training and evaluation data. - Class-conditional shift arises when the appearance of a class itself changes between domains. For example, tumor cells in early-stage cancer might look very different from those in late-stage cancer, even though they’re labeled the same, which can confuse models trained on just one type. Each of these domain shifts poses unique challenges, and addressing them requires tailored solutions. Our paper explores all four in depth, helping researchers recognize, diagnose, and respond to them with the right tools. To address domain shift, we explore domain generalization (DG), a way to build AI systems that can perform well even on new, unseen data from different settings. The paper provides a comprehensive overview of current DG methods, explains why traditional techniques often fail in real-world use, and evaluates the strengths and weaknesses of different strategies like data augmentation, stain normalization, meta-learning, and ensemble models. This paper also introduces clear guidelines and benchmarking results based on 28 advanced DG algorithms, helping researchers design better experiments and improve the reliability of AI systems in clinical practice. Although focused on computational pathology, the insights are relevant to a wide range of medical image analysis tasks.

Featured Image

Photo by Annie Spratt on Unsplash

Photo by Annie Spratt on Unsplash

Why is it important?

AI models in healthcare, particularly in pathology, can significantly support diagnosis and treatment planning. However, if these models don’t work consistently across different hospitals or patient groups, they become unreliable and potentially unsafe. This study addresses this crucial limitation by showing how to make AI models more generalizable. There are multiple challenges in transitioning from conventional pathology to AI-enabled digital pathology in practice, categorized under operational, technical, regulatory, ethical, and cultural challenges. We believe the domain generalizability of AI models plays an important role in resolving these hurdles in the path This paper is the first comprehensive survey focused specifically on domain generalization in digital pathology. The authors not only explain technical methods but also provide benchmarking experiments and practical guidelines. These contributions help bridge the gap between lab-developed models and real-world clinical applications, ultimately supporting safer, more equitable, and more widely usable AI in healthcare.

Perspectives