TLDR: This research paper highlights that AI models used in computational pathology, known as Foundation Models, are significantly affected by differences between digital slide scanners. This ‘scanner bias’ can lead to inconsistent predictions for the same tissue scanned on different devices, hindering clinical trust and adoption. To combat this, the authors propose ScanGen, a new contrastive loss function that helps these models learn to generalize better across various scanners. Experiments show ScanGen effectively reduces scanner bias while improving or maintaining the accuracy of predictions, such as for EGFR mutation in lung cancer, making AI tools more reliable for global clinical use.
Computational pathology (CPath) is transforming how medical professionals analyze tissue samples, especially through the use of Whole Slide Images (WSIs). Deep Learning (DL) models, particularly advanced Foundation Models (FMs), are central to this transformation, offering powerful insights. However, a significant challenge known as ‘scanner bias’ has emerged. This bias occurs because different commercially available slide scanners introduce subtle variations in image characteristics, such as sharpness, contrast, or artifacts, which can cause the same tissue specimen to appear differently depending on the scanner used. Consequently, a CPath model might produce varying outputs for the same tissue, undermining clinicians’ trust and slowing the adoption of these tools in real-world clinical practice.
The research paper, titled “Pathology Foundation Models are Scanner Sensitive: Benchmark and Mitigation with Contrastive ScanGen Loss,” delves into this critical issue. Authored by Gianluca Carloni, Biagio Brattoli, Seongho Keum, Jongchan Park, Taebum Lee, Chang Ho Ahn, and Sergio Pereira, the study provides empirical evidence that even modern pathology Foundation Models are susceptible to scanner bias. This is particularly concerning given the uneven global distribution of scanner types, which could lead to regional disparities in patient care if models are not robust across devices.
To address this problem, the authors propose an innovative solution called ScanGen. ScanGen is a contrastive loss function designed to mitigate scanner bias during the task-specific fine-tuning of FMs. In simple terms, ScanGen works by training the model to recognize that images of the same tissue specimen, even if acquired by different scanners, should be treated as highly similar. Conversely, it encourages the model to differentiate between images from different specimens, even if they were scanned on the same device. This process helps the model learn a more ‘scanner-agnostic’ representation of the tissue, enhancing its robustness to variations introduced by different scanning hardware.
The effectiveness of ScanGen was rigorously evaluated using a comprehensive multi-scanner dataset. This dataset included 323 specimens, each scanned by six different devices at 40x magnification and one scanner at 20x magnification, totaling 2,261 WSIs. The researchers applied their approach to the clinically relevant task of predicting Epidermal Growth Factor Receptor (EGFR) mutations in lung cancer from H&E-stained WSIs. They benchmarked five state-of-the-art FMs, including Phikon, UNI, Virchow, Gigapath, and H-Optimus-0.
The results were compelling. The study found that all tested FMs indeed suffered from scanner bias, as evidenced by visual clustering of images by scanner type rather than specimen type. However, with the integration of ScanGen, there was a notable improvement in generalization across scanners. Quantitatively, ScanGen significantly reduced the Coefficient of Variation (CoV) – a metric for scanner bias – across all tested FMs. Crucially, this improvement in scanner robustness did not come at the expense of predictive performance; in fact, the Area Under the Curve (AUC) for EGFR mutation prediction either improved or was maintained for each embedding type. ScanGen also demonstrated an ability to mitigate disagreements in predictions when slides were read at different magnifications.
Further analysis revealed that the more diverse the set of scanners used during ScanGen training, the better the generalization performance. This suggests that exposing the model to a wide range of scanner variations during training is key to achieving optimal robustness. The researchers also showed that ScanGen can be seamlessly integrated with other popular Multiple Instance Learning (MIL) techniques, consistently improving scanner agreement and maintaining or surpassing baseline AUCs.
Also Read:
- New Dataset and Framework Tackle Scanner Variability in Digital Pathology
- Enhancing Medical Image Analysis: A Hybrid Approach to Optimize Transfer and Self-Supervised Learning
In conclusion, this research highlights a critical, often overlooked, challenge in computational pathology: the sensitivity of AI models to digital slide scanners. By proposing and validating ScanGen, the authors offer a practical and cost-effective solution to enhance the robustness and generalization capabilities of existing pathology Foundation Models. This advancement is vital for building trust in CPath-based tools and facilitating their widespread, equitable deployment in clinical settings globally. For more details, you can refer to the full research paper available at arXiv:2507.22092.
