Virtual image generation of immunohistochemical stained tissue sections using an artificial neural network
Realistic immunohistochemical stained tissue section databases for training, certification, and machine standards.
Immunohistochemistry (IHC) is used to determine antigen distribution in a tissue and is widely used for diagnosis of cancers and other diseases. Diagnostic and research laboratories in the United States use locally devised IHC tissue slide preparation and scanning protocols to check scanner performance, assess student knowledge, and compare scanner quality. In addition, slides are used as part of competency practical exams for obtaining and renewing pathologist certifications. Evaluation of quantitative image analysis methods includes the use of an expert-generated reference standard and comparing computer results to that standard. However, due to inter- and intra-observer variability, the standards must be assessed and reviewed by multiple experts to be validated. This process is resource-intensive and limits the sample size for evaluation, drawing in the uncertainty inherent in a small sample size. To address such issues, sample sets have been created with histological image generation algorithms. However, the models are "unrealistic" and are not useable in validating analytical methods and do not match expectations of pathologist users.
Dr. Metin Gurcan and colleagues have developed a method of creating realistic, synthetic digital slide images of IHC stained tissues with known, fully-controlled ground truth using artificial neural networks. User-defined parameters enable image generation with: a) known number of positive/negative cells (count/percent); b) various cell types; c) mixed populations; d) altered morphology; e) varied spatial distributions including focal clustering/overlap; and f) various stains including different intensities and patterns of staining. Furthermore, these phantom digital histopathology tissue sections have background imaging that supports the realism of the sample, correcting a significant shortcoming of other approaches. For a given tissue type (breast, lung, prostate, lymphoma, etc.) the phantom tissue can be further complicated by introducing other objects such as lymphocytes, histiocytes, blood vessels, nerve bundles and fibrotic fibers. These in-silico standards will allow clinicians, research pathologists, and trainees to analyze accuracy, precision, and observer variability of IHC slides in a systematic, consistent manner. Unlimited training and testing samples can be generated to improve skillsets and algorithm development. During a test to assess "realism", synthetic images were generated and trials were conducted to assess the ability of experts to distinguish the real from phantom images. Experts were unable to detect differences in the two tissue specimen slides - overall, the accuracy in guessing real or generated was only 50% (33.3%-60%), similar to a random guess.