FDPNP10: Handling Whole Slide Image in Histopathology
Preface: I want to initiate a comprehensive series, “Fundamentals of Digital Pathology for Novice Pathologists (FDPNP)”. This endeavor aims to disseminate knowledge about the utilization of computer vision in the realm of modern pathology to pathologists who want to know more about digital pathology.
A whole slide image (WSI) is an image with a super-resolution. WSI is used as a digitalized substitution of the tissue slide, where pathologists investigate and make histopathologic diagnoses. However, the WSI is too large for a trivial graphical processing unit (GPU) to handle at a time.
Therefore, researchers have been investigating methods to effectively extract the information at the WSI level without consuming a lot of memory space [1]. These methods can be regarded as aggregation methods.
Image by Author
Aggregation Methods to Process WSI
According to [1], there are three categories of aggregation methods, including heuristic/statistical aggregation, data-driven aggregation, and clinically-driven aggregation.
Heuristic/statistical aggregation comprises global pooling approaches that do not rely on machine learning (parameters) or data at all. Examples include mean and max pooling.
Image by Author
Data-driven aggregation requires the inclusion of a set of learnable parameters. Examples of this method consist of multiple-instance learning (MIL) and hierarchical transformers. I mentioned MIL in more detail in a previous blog [2].
Image by Author
Clinically-driven aggregation extracts the WSI information in a way similar to pathologists by employing the microscopic methodology of the pathologist such as quantification of tumor-stroma ratio, tumor-infiltrating lymphocyte, and PD-L1 expression [3].
Image by Author
Discussion
Even when applying these aggregation methods, it’s worth noting three significant points of discussion regarding the handling of WSI in histopathology.
Firstly, effectively extracting relevant small patches from WSIs is not a feasible task. This is primarily because most WSIs, especially those from biopsy specimens, contain numerous irrelevant structures, including space, normal tissue, and blood. While these structures are distinguishable to a pathologist, they can pose challenges when training AI models.
Secondly, WSIs contain a multitude of histopathologic and technical artifacts. These can include foreign objects, tissue cracking, ink, or stain fading. These artifacts may degrade the quality of training/testing WSIs, and thus, aggregation methods should account for them.
Image by Author
Finally, there is considerable visual variation between institutions, scanners, staining protocols, and even technicians. Each institution has its own protocols for tissue storage, logistics, and processing. Differences in color, scaling, and image quality occur between different scanners. The intensity of the stains, including hematoxylin and eosin, is highly sensitive to the staining protocol and the technicians who perform the tissue staining.
Image by Author
Conclusion
Although computer vision models have made a tremendous change in the field of pathology, processing and handling a WSI is not an easy task. The way we extract the WSI information directly affects how effectively the model operates.
To those who may be interested, I hope you enjoy this mini-review.
References
-
Bilal, Mohsin, et al. “An Aggregation of Aggregation Methods in Computational Pathology.” Medical Image Analysis, vol. 88, 1 Aug. 2023, pp. 102885–102885, https://doi.org/10.1016/j.media.2023.102885. Accessed 4 Feb. 2024.
-
Le, Minh-Khang. “Multiple Instance Learning (MIL) and Its Utility in Whole Slide Image (WSI) Analyses.” Medium, 3 Oct. 2023, medium.com/@minhkhangle.phd/multiple-instance-learning-mil-and-its-utility-in-whole-slide-image-wsi-analyses-3acb67f5434b.
-
Le, Minh-Khang. “Histopathological Biomarker Extraction by Computer Vision.” Medium, 23 Nov. 2023, medium.com/@minhkhangle.phd/histopathological-biomarker-extraction-by-computer-vision-720ffb099605. Accessed 8 Mar. 2024.