to BioTechniques free email alert service to receive content updates.
Use and validation of epithelial recognition and fields of view algorithms on virtual slides to guide TMA construction
Sanford H. Barsky1,2, Lynda Gentchev3, Amitabha S. Basu4, Rafael E. Jimenez3, Kamel Boussaid4, and Abhi S. Gholap4
1Department of Pathology, University of Nevada School of Medicine, Reno, NV, USA
2Nevada Cancer Institute, Las Vegas, NV, USA
3Ohio State University College of Medicine, Columbus, OH, USA
4BioImagene, Inc., Cupertino, CA, USA
BioTechniques, Vol. 47, No. 5, November 2009, pp. 927–938
Full Text (PDF)
Supplementary Material

While tissue microarrays (TMAs) are a form of high-throughput screening, they presently still require manual construction and interpretation. Because of predicted increasing demand for TMAs, we investigated whether their construction could be automated. We created both epithelial recognition algorithms (ERAs) and field of view (FOV) algorithms that could analyze virtual slides and select the areas of highest cancer cell density in the tissue block for coring (algorithmic TMA) and compared these to the cores manually selected (manual TMA) from the same tissue blocks. We also constructed TMAs with TMAker, a robot guided by these algorithms (robotic TMA). We compared each of these TMAs to each other. Our imaging algorithms produced a grid of hundreds of FOVs, identified cancer cells in a stroma background and calculated the epithelial percentage (cancer cell density) in each FOV. Those with the highest percentages guided core selection and TMA construction. Algorithmic TMA and robotic TMA were overall ~50% greater in cancer cell density compared with Manual TMA. These observations held for breast, colon, and lung cancer TMAs. Our digital image algorithms were effective in automating TMA construction.


The high-throughput screening (HTS) revolution has taken center stage in cancer research: cDNA and oligo-spotted microarrays, proteomic blots, and tissue microarrays (TMAs) have become standard research tools (1,2,3). Of these three methodologies, TMAs represent an anachronism since both their construction and interpretation are manual and subjective rather than automated and truly high-throughput (4). Paradoxically, the need for reliable and efficient TMA creation and interpretation has increased for scientists discovering and validating putative cancer biomarkers (5). The final and ultimate proof of any putative cancer biomarker discovered by genomic or proteomic approaches lies in an evaluation of its pattern of in situ expression that only TMAs can provide. With the advent of digital pathology approaches and imaging technology that can scan whole microscopic slides into virtual images with high resolution (6), we wondered whether mathematical insights into the morphology of cancer cell images in virtual slides might translate into imaging algorithms that could effectively guide the selection of the optimal areas of tissue for coring and the construction of TMAs. Because of the increasing demand for TMAs predicted to occur over the next decade and their expanding role in both biomarker discovery validation in cancer research (7)—and because one key rate-limiting step in the use of TMAs for this purpose is the time and effort it takes for their manual construction—we felt it would be highly desirable to automate their construction. Because we wanted to automate the construction of TMAs, we needed to create algorithms that could select cores with the highest proportion of epithelial cells (i.e., cancer cells) and that these algorithms needed to be at least as good as the manual selection of cores. After these algorithms were devised and evaluated, we were also able to construct a robot to automate the construction of TMAs under algorithmic guidance.

Materials and methods

This study was reviewed by The Ohio State University College of Medicine Institutional Review Board (IRB) and approved (2006C0042).

Selection of cases

In this study, we retrieved 300 cases (consisting of two whole slides each and two corresponding paraffin blocks) of human breast cancer (100 cases), colon cancer (100 cases), and lung cancer (100 cases), which were used to make the different TMAs. Prior to TMA construction, each whole slide was scanned and made into a “virtual slide” and analyzed algorithmically. The algorithmic analysis consisted of applying a virtual “grid” to the image and dividing the slide into fields of view (FOVs) and subsequently analyzing each FOV with epithelial recognition algorithms (ERAs) designed to quantitate the epithelial cell (cancer cell) percentages.

Image acquisition

Image acquisition utilized a scanner, either the Aperio ScanScope T2 System (Aperio, Vista, CA, USA) or the iSCAN System (BioImagene, Inc., Cupertino, CA, USA), which was capable of producing images with a resolution of 20 pixels/10 µm. Following image acquisition, images were screened for quality, enhanced and processed. For image acquisition, any commercially or freely available imaging system utilizing either a scanner or a microscope with attached digital or analog camera can be used, as long as the images are produced with a resolution of 20 pixels/10 µm.

Software for image analysis

For our overall approach of writing the software for image processing and analysis, we used Visual C++. With this software language we also created novel ERAs and FOV algorithms that recognize cancer cells and identify the areas of the slide with the highest density of cancer cells. The steps of the FOV and ERA algorithms (Supplementary Tables S1 and S2), as well as more-detailed flowcharts and descriptions of the algorithms (Supplementary Tables S3–S7) are provided.

Gridding of the virtual slide by the FOV algorithm

Each virtual slide was divided into a grid of squares (termed FOVs), each measuring 1 mm × 1 mm or 2000 pixels × 2000 pixels (Figure 1). The tissue slices appearing on glass slides exhibited a wide range of sizes but each could invariably be overlaid by a grid. The FOVs within the grid were individually analyzed (Supplementary Table S1). This approach ensured that the dimensions of each FOV were less than or equal to the predefined dimensions for FOV (2000 × 2000 pixels). On a given slide there could be several pieces of tissue each termed an area of interest or identification (AOI). In that case, each AOI was gridded and divided into FOVs of equal dimensions. Some AOIs, upon gridding, contained FOVs with little or no tissue, since the grid overlay was a collection of squares and the tissue fragment border usually irregular. At a designated threshold for the ratio of tissue to no tissue in a given FOV, that FOV could be blocked from further analysis. Alternatively, separate AOIs could be virtually merged before gridding was performed; where appropriate, blocked FOVs could then be unblocked if the merged image produced a greater ratio of tissue to no tissue.

  1    2    3    4    5