Color correction technology for histopathological image datasets could support machine learning diagnostic tools

In histopathology, where tissues are examined under the microscope to understand and diagnose disease, dyes are an essential tool. Simply put, dyes are carefully selected or manufactured chemicals that adhere to specific cellular components. When viewed under the microscope, they help the user to more easily distinguish cellular structures by changing the colors observed.

Datasets with colorized images showing both normal and diseased tissue are valuable for training machine learning models that can help doctors evaluate difficult cases and mitigate personal biases in diagnosis. To ensure that these models work properly, it is important to minimize color differences between the images used for training and those they analyze in real-world scenarios. Using so-called “domain adaptation” techniques, color variations resulting from the unique experimental setups in different labs can be corrected, creating more consistent and comparable data.

In a recently published study Journal of Medical ImagingResearchers at the University of North Carolina at Chapel Hill (USA) have proposed a novel domain adaptation technique. The proposed method, called Stain Simultaneous Augmentation and Normalization (Stain SAN), can help make stained histopathological image datasets more useful for many new machine learning-based classification systems, ultimately leading to improved diagnostic tools.

Stain SAN involves three main steps: stain extraction, color matching, and intensity matching. In the first step, the original stain image is decomposed into a product of two matrices: one containing color information and the other containing light intensity information for each pixel. In the matching step, the color distribution in the color matrix is ​​modified by a statistical process that takes into account the training images and guarantees that the modified colors are within a target distribution. Finally, in the third step and before image reconstruction, the intensity matrix is ​​subjected to random perturbation. This helps to increase the diversity of possible stain regions.

“The main advantage of Stain SAN is that it combines the strengths of previous stain matching methods while overcoming their inherent weaknesses,” explains Dr. Taebin Kim, the lead researcher. “Other established techniques, including stain normalization, stain enhancement, and stain blending, can be considered special cases of Stain SAN.”

The researchers tested their approach both qualitatively and quantitatively on histopathology images from publicly available datasets. Based on their observations as well as feedback from an experienced pathologist, the researchers found that image datasets processed with Stain SAN resulted in more uniform colors with better generalized areas of staining. In addition, Stain SAN increased the contrast between the nucleus and cytoplasm in each cell and highlighted differences between tumor cells and supporting tissue.

The team also trained machine learning-based classifiers on datasets processed using different domain adaptation techniques and tested their performance on processed images from another dataset. Interestingly, Stain SAN outperformed all the previously mentioned methods, delivering significantly higher accuracy.

“Our results clearly demonstrate the improvements made over the course of developing these methods, culminating in a substantial improvement by Stain SAN,” comments Kim. “In addition, we demonstrated that Stain SAN achieves results comparable to a state-of-the-art deep learning-based approach without requiring separate training for staining adaptation or access to the target domain during training, which would be unrealistic in clinical practice. This underscores the effectiveness and computational power of Stain SAN.”

Developing efficient domain adaptation techniques such as Stain SAN is essential to bridge the gap between machine learning systems and their applications in healthcare. The research team is already planning possible improvements to their method and conducting further tests on other datasets.

“Our findings confirm Stain SAN as a robust approach for tailoring staining domains in histopathological images, with implications for the further development of computational tasks in this field,” concludes Kim, looking optimistically to the future.

These efforts will pave the way for more precise and practical diagnostic protocols, saving time for both physicians and patients.