Abstract
Our results show that estimation of tissue epithelium percentage using our color-based segmentation method correlates well with pathologists' estimation of tissue epithelium percentages. The epithelium contents estimated by color-based segmentation may be useful in immuno-based analysis or clinical proteomic analysis of tumor proteins. The codes used for epithelium estimation as well as the micrographs with estimated epithelium content are available online.
Introduction
In the search for cancer biomarkers, researchers frequently utilize high-throughput quantitative protein assays, such as ELISA or mass spectrometry, on homogenized tissue specimens. However, these "bulk" measurements are often confounded by tissue heterogeneity. Tumors typically possess a significantly higher proportion of epithelial cells—the primary site of carcinoma origin—compared to healthy control tissues, which are often richer in stroma.
Consequently, an observed increase in protein concentration in a tumor sample may reflect simple changes in tissue composition rather than true cellular overexpression. Accurate interpretation of these assays requires normalization based on the ratio of epithelium to stroma. Traditionally, this ratio is estimated manually by pathologists, a process that is labor-intensive and prone to inter-observer variability.
Methods
The study developed an automated pipeline to quantify tissue composition using 44 prostate tissue specimens (36 tumor, 8 normal) from the Johns Hopkins University School of Medicine.
Image Segmentation and Classification
Histology slides stained with hematoxylin and eosin (H&E) were digitized. A k-means clustering algorithm, implemented in MATLAB, was used to segment pixels into four distinct color categories: white (lumen/empty space), bright pink (stroma), dark pink (stroma), and purple (epithelium).
To classify tissue regions, each image was divided into a $20 \times 20$ grid. The algorithm calculated color ratios within each grid cell to determine the predominant tissue type. The dataset was split equally for training and validation against manual markings.
Protein Normalization
Concentrations of two proteins, EpCAM and Cathepsin L (CTSL), were measured via ELISA. These values were then normalized using the calculated epithelium percentage to determine "protein per unit of epithelium," providing a more accurate biological comparison between normal and malignant tissues.
Results
The automated segmentation method achieved an accuracy of 84.41%, with balanced false positive and false negative rates (~8%). The computer-estimated epithelium percentages showed a strong correlation with manual markings ($R^2 = 0.965$) and a Pearson correlation coefficient of 0.72 compared to independent pathologist estimates.
Impact on Biomarker Interpretation
Normalization significantly altered the interpretation of protein data:
- EpCAM: Remained significantly higher in tumors after normalization ($p < 0.001$), confirming its validity as a biomarker. However, the apparent 4.75-fold increase in raw data was corrected to a more realistic 2-fold biological increase.
- CTSL: While raw data suggested a significant increase in tumors ($p < 0.001$), this difference disappeared after normalization. This indicates that the perceived upregulation of CTSL was an artifact of increased epithelial density in tumor samples, resolving previous contradictions between bulk ELISA data and immunohistochemistry (IHC) findings.
Discussion
This study demonstrates that automated color-based segmentation can effectively replace or augment manual pathological review for tissue normalization. By accounting for tissue heterogeneity, researchers can distinguish between true biological overexpression and compositional shifts.
Future work should address potential variability in H&E staining across different laboratories, which may necessitate recalibration of the k-means algorithm. Additionally, ensuring that the histology slice used for estimation is representative of the bulk tissue mass remains critical for accurate normalization. This technique holds promise for application across other "omics" platforms, including mRNA expression and protein activity assays.
Summary updated