A deep learning method for automatic evaluation of diagnostic information from multi-stained histopathological images

Manual screening of large-scale histopathological images is an extremely time-consuming, laborious and subjective procedure. Accurate evaluation of diagnostic information from multi-color stained images requires expertise due to the complex nature of histopathology and the lack of quantifiable measurement. In this work, a novel deep learning method is developed based on a convolutional siamese network, in which the information quantification task is transformed into a similarity assessment between lesion and non-lesion patterns on histopathological images. The subtle changes underlying the microstructure of tissue biopsies can be captured through an optimization of training loss within a low-to-high-level feature space. A new information score is introduced to quantify the abnormality in tissue appearance and stain pattern. Experiments on 3 independent data cohorts including 5 types of color-stained images demonstrate that our method can achieve promising performance compared with state-of-the-art methods. Results show that the proposed information score can serve as an effective measure to evaluate the importance of multi-stained images, and ultimately facilitate automatic diagnosis for clinical multi-stained histopathology. •An automatic method is presented for quantifying multi-stained histopathological images.•A new information score is computed to evaluate diagnostic information.•The classification provides supplementary opinion in the cross-slide pathology diagnosis.•The method can serve as a generalized framework to analyze different histology subtypes.