Author: Diana Cardona, Casey C. Heirman, William Jeck, Kyle J. Lafata, Xiang Li, Lauren M. Neldner, Jeffrey S. Nelson, Megan K. Russ, Ehsan Samei π¨βπ¬
Affiliation: Duke University, Department of Radiation Oncology, Duke University, Department of Pathology, Duke University, Clinical Imaging Physics Group, Department of Radiology, Duke University Health System π
Purpose: Medical physicists have traditionally supported radiation-based medicine, but their expertise can translate to other image-based fields including pathology. As pathology transitions to digital practices, physicistsβ proficiency in image analysis can be put toward optimization of slide imaging systems and workflow consistency. Variability due to slide preparation, imaging, post-processing, and display remains a critical area of study. This study explores methods to quantify in image chromaticity and contrast arising from slides scanned on two whole slide imaging scanners used interchangeably within one department.
Methods: Five H&E-stained mouse buccal tissue slides were imaged using two Leica Aperio scanners: a GT450 and AT2. A machine learning algorithm trained to detect lymphocytes segmented all lymphocytes in each image. The chromaticity of each lymphocyte and its contrast compared to a halo region around its segmentation boundary were evaluated. Chromaticity was measured by sampling RGB pixel values of segmented lymphocytes and converting to CIE1976 coordinates. Contrast was calculated as difference between lymphocyte and halo CIELUV coordinates, expressed in units of just noticeable differences (JND).
Results: Images of the same slide differed markedly between scanners, with lymphocyte chromaticity differences exceeding the perceivability threshold (Ξuβvβ = 0.065 > 0.04). Compared to the GT450, the AT2 produced images with substantially higher lymphocyte contrast. The machine learning algorithm detected 46% more lymphocytes in GT450 images on average, many of which were determined to be false positives.
Conclusion: Differences in lymphocyte chromaticity and contrast between scanner models underscore the need for optimization in digital pathology to ensure consistent image quality for pathologists and AI-based tools. Future work will investigate methods of mitigating these differences, as well as studying the impact differences in lymphocyte conspicuity have on diagnostic accuracy. Medical physicists can play a pivotal role in applying quantitative image quality expertise to support reliable diagnostics in this evolving field.