Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Bioinformatics, Computational and Systems Biology

Bioinformatics, Computational and Systems Biology - Poster Session B

Poster A7 - Development of Method for Quantitative Assessment of Digital Pathology in Woodchuck HCC Model Using Image Segmentation and Image Processing

Thursday, October 24, 2024

2:45 PM – 3:45 PM EST

Location: Exhibit Hall E, F & G

Presenting Author(s)

Joshua Cheuk (he/him/his)

Biomedical Engineering Summer Internship Program Scholar
National Institute of Biomedical Imaging and Bioengineering, United States

Co-Author(s)

NV

Nicole Varble, PhD

Senior Scientist
Philips, United States
ML

Ming Li, PhD (she/her/hers)

Senior Scientist
National Institutes of Health, Maryland, United States
Andrew S. Mikhail, PhD

Staff Scientist
Center for Interventional Oncology, Radiology and Imaging Sciences, Clinical Center, National Institutes of Health, United States
WP

William F. Pritchard, MD PhD

Physician
Center for Interventional Oncology, Radiology and Imaging Sciences, Clinical Center, National Institutes of Health
Bethesda, Maryland, United States
JK

John W. Karanian, PhD

Medical Physiologist
Center for Interventional Oncology, Radiology and Imaging Sciences, Clinical Center, National Institutes of Health, Maryland, United States
LS

Laetitia Saccenti

Research Fellow
National Institutes of Health, United States

Primary Investigator(s)

Bradford J. Wood, MD (he/him/his)

Director, NIH Center for Interventional Oncology & Chief IR
National Institutes of Health
Bethesda, Maryland, United States

Co-Author(s)

Robert Morhard, PhD (he/him/his)

Postdoctoral Fellow
Center for Interventional Oncology, Radiology and Imaging Sciences, Clinical Center, National Institutes of Health
Bethesda, Maryland, United States

Introduction: Hepatocellular carcinoma (HCC), is the fifth most common cause of cancer worldwide as of 2023. HCC is a malignant tumor of the hepatocytes. Common causes of HCC include hepatitis B or C infection, heavy alcohol use, cigarette smoking, and hereditary hemochromatosis. Common treatment methods include tumor ablation, liver transplantation, surgical resection, systemic chemotherapy, and transarterial therapies.

Diagnosis and treatment of HCC can be done via imaging studies, such as computed tomography (CT), and via biopsy, after which histopathological analysis can be performed. The histopathological examination requires analysis by a pathologist, making it difficult to communicate findings to more junior individuals and medical and research personnel who aren’t pathologists. While metrics such as the Hounsfield unit (HU) used in CT describe the physical density of the tissue, such metrics are not currently used in histopathology. Thus, a standardized metric must be created to analyze the tumor, surrounding tissue, and the cellular anomalies found in Hematoxylin and Eosin (H&E) histopathology slides.

Materials and

Methods: Histopathology samples were extracted from woodchucks chronically infected with woodchuck hepatitis virus that developed hepatocellular carcinoma (HCC). Liver and tumor tissues were biopsied, sectioned, stained with hematoxylin and eosin (H&E), and digitized.
For image processing, the 85,679x80,708 pixel histopathology slides were divided into thousands of 256x256 pixel tiles, facilitating easier manual and computational analysis. These tiles were manually sorted into three categories: those partially covered, fully covered, and fully covered with cells plus a blood vessel. The average pixel value for each tile category was calculated.
To enhance the viability of hematoxylin (H) and eosin (E), normalization of pixel values was performed using Python 3.12.4, CV2, Sci-kit image, Scipy, OpenSlide, and Pillow. The normalized tiles were split into separate H and E images, improving cell structure differentiation for machine and human cell counting.
For image segmentation, nuclei within the tiles were segmented, counted, and their areas analyzed. The segmentation model's accuracy was tested using three types of tiles: those partially covered with cells but had enough to be considered a fully covered tile, those fully covered with cells, and those fully covered with cells plus a blood vessel. Manual cell counts from 3 readers established a baseline for comparison with automated counts. Image segmentation and analysis were conducted using Python with CSBDeep, StarDist, and ScikitImage.
This approach ensured accurate extraction, processing, and analysis of histopathology samples. The scale is 120% of its original size. The probability threshold and the Non-maximum Suppression were at default settings.

Results, Conclusions, and Discussions: Out of the 26,378 tiles created, 292 tiles were selected as a sample to be normalized. 77 were blank, 108 were fully covered with cells, and 107 were partially covered with cells. This was used to represent a small sample of the thousands of tiles. Using the average pixel value of the good, fully covered tiles (164.1 + 48.3), only the tiles within this range were selected to be normalized using Openslide and had H separated from E.

The SCT had an average of 50 + 4.24 cells, the FCT had a mean of 98.67 + 9.87, and the FCT-BV had a mean of 73.33 + 1.53.

The percent difference was the lowest when comparing the manual average (MA) to H-only tiles. FCT of MA to H-only was a 14.89% difference, SCT to H-only had a percent difference of 8.33%, and FCT-BV tile to H-only had a percent difference of 47.88%. Of the three tile types, the SCC is the most accurate.

The most accurate normalized and H-only tiles were SCC with a percent difference of 12.77% and 8.33% respectively. This could mean that the image segmentation model had an easier time identifying the shapes of the cells, as the non-normalized eosin makes an obvious cell outline around the nuclei, as these cell types have easily identifiable nuclei with a partially white background.

The least accurate original, normalized, and H-only tile was FCT-BV. It has the lowest percentages for original, normalized, and H-only with 52.14%, 54.33%, and 47.88% respectively. Due to the abnormal shapes and the overlapping cells of the endothelial cells, the StarDist image segmentation model has a hard time identifying the cells on original, normalized, and H-only tiles. Thus, StarDist image segmentation may not work best on endothelial cells.
One overall limitation is that StarDist has its own normalization model, so using Openslide to normalize the images may have made it difficult for StarDist to identify the cells.

Future directions for this project include analyzing the abnormal structures in each tile to identify nucleic and cellular anomalies and helping pathologists diagnose cancer quickly.

Acknowledgements (Optional):