Top Accurate Cell Detection and Segmentation Using A Pen, Agile Digital Image Processing for Deep Learning Models

Abstract Introduction/Objective The Hematopathology labs suffer from insufficient productivity, profitability and increased liability. The predicted shortage in the Hematopathologists in the USA is 25% by 2040. Medical errors and results variability due to fatigue and stress could also increase. At least 99% specificity or 99% sensitivity is required in automated visual review of peripheral blood and bone marrow smears to be generalizable and work-fit. Fortunately, the growing digital Hematopathology scanners market besides the vastly developed artificial intelligence (AI) and machine deep learning (DL) technologies could make this realized soon. The AI/DL production cycle includes (1) images collection, (2) image processing/data science, (3) technology, (4) engineering, (5) testing and validation. Methods/Case Report Integration with a Hematopathology lab using digital scanners and/or integrated cameras on microscopes is the mean for whole slide images (WSI) collection. Image processing refers to AI/DL spoon-feeding with accurate, diverse and highly-dissimilar materials for learning. This requires providing hundreds of thousands examples of (a) WSI context correction for stain and resolution normalization, (b) cellular dimensions on the WSI patches for detection, (c) cellular segmentation/boundaries determination for counting (d) cellular labels for classification. Comparing with the confirmatory testing results including flow cytometry and immunohistochemistry is essential in both image processing and validation. Currently, image processing resembles 80-90% of the AI/DL cycle/pipeline which may take years of hectic work. Our proposed developed tool enables the Hematopathologists to use pens and draw on the boundaries of each cell in a field/patch. 3-15 seconds are needed for each cell. The required ground truth/masks for segmentation and dimensions for detection are generated in real-time with no need for data scientists. Results (if a Case Study enter NA) NA Conclusion Similar to the automated agile development-operation loop concept (DevOps), enabling the Hematopathologists to perform image processing directly will reduce the production cycles from years to weeks. No technology can perform qualitatively better than the human eyes, the proposed pen-based cell detection and segmentation tool results in 10-50 fold increase in quality and quantity. Promisingly, there is a trend among the AI/DL platforms companies to automate the AI/DL coding.