Automated inmunohistochemical staining quantification in human biopsies: preliminary results using deep learning

Abstract

Among the current challenges in histopathological assessment for diagnosis in clinical contexts is an accurate determination of the actual tissue malignancy. This task is often performed using microscopy over immunohistochemical (IHQ) staining applied on tissue samples, on which several specialists judge the tissue con- dition following specific criteria. However, this task is proven to be prone to high inter- and intra-subject variance, which raises the need to elaborate more robust tools and frameworks to assist on this task. The recent influx of deep learning technologies, which are proven to be successful in a variety of contexts, appears to be an adequate alternative in this context. In this aim, we present a joint effort between research groups from Cancer Biology Laboratory (INIBIBB-CONICET) and the Imaging Sciences Laboratory (LCI- UNS-CONICET). Starting with IHQ stained images taken with Olym- pus CX31 microscope from thyroid and breast cancer biopsies, we applied a Mask C-RNN network for cell nuclei detection. For this purpose, we retrained the net with a series of labeled examples pro- vided by the biochemical specialists. After this initial detection, a ROI was determined surrounding the nuclei, within which the proportion of diaminobenzidine stain (brown-colored precipitation) is computed as a proxy indicator of the Immunoreactive Score (IRS). For this, a Random Forest classifier was trained using stain/no stain labeled pixels also provided by the experts. The results appear promising in the sense that the resulting system is able to consistently provide malignancy assessment even in difficult cases or when the quality of the microscopy acquisition is below standard.