Tissue pattern recognition error rates and tumor heterogeneity in gastric cancer

The anatomic pathology discipline is slowly moving toward a digital workflow, where pathologists will evaluate whole-slide images on a computer monitor rather than glass slides through a microscope. One of the driving factors in this workflow is computer-assisted scoring, which depends on ap- propriate selection of regions of interest. With advances in tissue pattern recognition techniques, a more precise region of the tissue can be evaluated, no longer bound by the pathologist's patience in manually outlining target tissue areas. Pathologists use entire tissues from which to determine a score in a region of interest when making manual immunohistochemistry assess- ments. Tissue pattern recognition theoretically offers this same advantage; however, error rates exist in any tissue pattern rec- ognition program, and these error rates contribute to errors in the overall score. To provide a real-world example of tissue pattern recognition, 11 HER2-stained upper gastrointestinal malignancies with high heterogeneity were evaluated. HER2 scoring of gastric cancer was chosen due to its increasing im- portance in gastrointestinal disease. A method is introduced for quantifying the error rates of tissue pattern recognition. The trade-off between fully sampling tumor with a given tissue pat- tern recognition error rate versus randomly sampling a limited number of fields of view with higher target accuracy was mod- eled with a Monte-Carlo simulation. Under most scenarios, stereological methods of sampling-limited fields of view out- performed whole-slide tissue pattern recognition approaches for accurate immunohistochemistry analysis. The importance of educating pathologists in the use of statistical sampling is dis- cussed, along with the emerging role of hybrid whole-tissue imaging and stereological approaches. BACKGROUND Despite recent improvements in surgical techniques and chemotherapy treatments, locally advanced/meta- static gastroesophageal junction (GEJ) and gastric cancer (GC) are still associated with poor clinical outcome. Re- sults from the Trastuzumab for Gastric Cancer trial demonstrated that trastuzumab, a monoclonal antibody directed against the extracellular domain of HER2, can enhance the efficacy of cytotoxic chemotherapy in pa- tients with advanced gastric and gastroesophageal can- cers. The pathologic analysis for HER2 in GC by immunohistochemistry (IHC) is currently performed with either manual scoring following the Hofmann-modified scoring criterion for GC 1 or the use of the Food and Drug Administration (FDA)-cleared HER2 algorithms de- signed for breast cancer. 2,3 To date, there are no image analysis algorithms that have been developed or have been approved specifically for the evaluation of HER2 protein overexpression in GC. Tumor heterogeneity presents a difficult problem in