A mixture of experts (MoE) model to improve AI-based computational pathology prediction performance under variable
Yujie Xiang1, Bojing Liu1, Mattias Rantalainen2
1Department of Medical Epidemiology and Biostatistics, Karolinska Institutet, Stockholm, Sweden.
Background:
AI-based models for analysis of histopathology whole slide images (WSIs) are now common. However, image quality, particularly unsharp areas of WSIs, impacts model performance. In this study we investigate the impact of blur on deep learning models for WSI analysis. We propose a mixture of experts (MoE) strategy that mitigates the impact of unsharp areas in WSIs on classification performance by combining predictions from multiple expert models trained on data with varying levels of blur.
Methods:
The study included hematoxylin and eosin (H&E) stained WSIs from 2093 breast cancer patients. Classification of histological grades 1 and 3 was used as a primary benchmarking case, and prediction of immunohistochemistry (IHC) markers (ER, PR, HER2) from H&E as a secondary case. The proposed MoE strategy to improve robustness against blur was evaluated in both a deep CNN model (CNN_CLAM and MoE-CNN_CLAM) and a Vision Transformer-based histopathology foundation model (UNI_CLAM and MoE-UNI_CLAM). For each architecture, a baseline model was trained on sharp images, and multiple expert models were trained on tiles with added Gaussian blur at different levels. Model performance (area under the ROC curve) was evaluated under multiple levels of uniform blur, as well as in several simulated scenarios with a mixture of blur levels within the WSIs.
Results:
Baseline model performance degraded with increasing blur for all evaluated architectures. Individual expert models trained on data with simulated Gaussian blur performed better on unsharp images compared to baseline models. The proposed MoE consistently outperformed its respective baseline models in simulation scenarios with various degrees of blur within WSIs. MoE-CNN_CLAM outperformed the baseline CNN_CLAM under moderate (AUC: 0.868 vs. 0.702) and mixed blur conditions (AUC: 0.890 vs. 0.875). MoE-UNI_CLAM outperformed the baseline UNI_CLAM model in both moderate (AUC: 0.950 vs. 0.928) and mixed blur conditions (AUC: 0.944 vs. 0.931).
Conclusions:
Unsharp image areas are common in WSIs and impact prediction performance. The proposed MoE strategy provided equal or substantially improved prediction performance under all evaluated test scenarios. The proposed methodology has the potential to increase quality and reliability of AI-based pathology models in both research and clinical applications.
