Effects of Color Calibration via ICC Profile on Inter-scanner Generalization of AI Models

USCAP 2022 Annual Meeting, Tempus-authored
Authors Kshitij Ingale, Rohan Joshi, Irvin Ho, Aicha BenTaieb, Martin Stumpe


Generalizing deep learning models to whole slide images (WSIs) from different sources is a well-known challenge in digital pathology. Scanners used for digitization of slides usually yield images in a different color space than the standard RGB (sRGB) space, characterized by the scanner ICC profile. Some scanners do not apply color space transform to sRGB during WSI creation, limiting the utility of imaging algorithms. Data augmentation is a popular method to aid model generalization, however, it requires retraining a deployed model as well as additional experiments to tune the extent of color augmentation. In this study, we evaluate the relative effect of color augmentation and ICC profile transformation (which does not require re-training) by assessing the performance of a model trained to predict microsatellite instability from H&E slides with and without these procedures.


For this study, a dataset of 1203 slides scanned with both Philips and Aperio scanners was curated and split into training (962 slides) and holdout-test sets (241 slides). Models were trained with or without color augmentation using only Philips WSIs, and evaluated on an independent Phillips and Aperio hold-out test set with (or without) ICC profile application. This process was repeated for 2 variants of attention based multiple instance learning model architectures (termed model 1 and model 2).


Figure 1a-c demonstrates color differences in Philips and Aperio WSIs with and without ICC. Figure 1d highlights that the ICC sRGB transform partially corrects differences in saturation for the Aperio scanner and moves the mean closer to the Philips mean. Model performance on unseen scanner slides is lower without color augmentation or ICC profile transformation (Figure 2). This gap can be bridged by retraining models with color augmentation. Additionally, ICC profile transform either improved model generalization without requiring retraining (model 1, Figure 2a) or produced qualitatively similar results(model 2). ICC profile transform does not negatively impact performance when used with or without color augmentation.


Using ICC profile transform can yield similar or better results on unseen scanner slides without model retraining. Color normalization via ICC profile could facilitate inter-scanner generalization when data is scarce. It can also be used as a part of the standardization workflow in digital pathology.