ABSTRACT Deep learning networks trained for classification of single cells in hematoxylin and eosin‐stained (H&E) sections are limited in accuracy. We hypothesize the accuracy of a pathologist to be 98%–100%. Digital classification should be on par with this to be relevant. Here is presented a human‐in‐the‐loop approach, where cell classification is fine‐tuned by the pathologist during microscopy. The approach was established and tested on H&E‐stained histological slides of pulmonary adenocarcinomas (LUADs). Networks pretrained on a limited selection of typical cells reached an average true positive rate (TPR) of 0.989 for classification of LUAD tumor cells within a high‐power field (HPF), with 20 additional tumor cells included from the tumor investigated. Results were consistent on three HPFs, yielding means of Dice = 0.991, TPR = 0.987, precision = 0.996, and specificity = 0.988. Networks trained and tested on a smaller dataset of LUADs from The Cancer Genome Atlas demonstrated means of Dice = 0.951, TPR = 0.944, precision = 0.96, and specificity = 0.951. Classification of LUAD tumor cells by networks based and tested on data from the same source is comparable to that of a pathologist. The output is interpretable, validatable, and applicable in routine pathology.
Stampe et al. (Wed,) studied this question.