Confidence Estimation in Stem Cell Classification

We study the problem of supervised classification of stem cell colonies and confidence estimation of the attained classification labels. The problem is investigated in the application context of heterogeneity labels of stem cell colonies observed by using fluorescent microscopy imaging. Given the features of colonies using numerous image statistics, we report the classification results using adaptive k-Nearest Neighbor (NN) algorithm. This algorithm minimizes typical k-NN classification bias by giving more weight to more informative features in predicting class posterior probabilities. We then estimate the confidence of each prediction for unlabeled data using transductive p-value and strangeness metrics. We show that such an introspection can gradually increase the accuracy of learned model, quantify false positives, and guide the resource-limited manual colony annotation process to provide training labels for the less confident unlabeled samples.