Quantitative genetic‐learning pattern recognition image analysis in cancer tissue biobank quality assurance

Quality assurance is needed to exploit the full potential of tissue biobanks. Histological evaluation of banked specimens serves to confirm diagnoses and provide a surrogate assessment of tissue features, including tumor presence, in biobank specimens. Histology pattern recognition software was tested as a means for automated, reproducible tissue feature quantification. Unique algorithms were optimized to evaluate spontaneous lymphomas (n=70), osteosarcomas (n=71), and melanomas (n=80) from the Canine Comparative Oncology and Genomics Consortium. Optimization focused on developing histology‐specific algorithms by training on known specimens until specificity and sensitivity approached 95%. Analysis of unknowns provided estimates of percent tissue section area occupied by stroma, tumor, and necrosis, the latter when subjectively considered >20%. Expert review revealed inappropriate tissue feature segmentation in a limited number of sections, which varied according to tumor type. > 90% of the area occupied by tumor was accurately annotated for lymphomas, while 25/80 melanomas included miscalls, primarily due to the marked pleomorphism. Pattern recognition software provides reproducible quantitative information for biobank quality assurance; however, pathologist oversight is important, especially for pleomorphic cancers. Support: NCI Intramural Research. CCOGC support from Pfizer.