Impact of stain normalization and patch selection on the performance of convolutional neural networks in histological breast and prostate cancer classification

Background Recently, deep learning has rapidly become the methodology of choice in digital pathology image analysis. However, due to the current challenges of digital pathology (color stain variability, large images, etc.), specific pre-processing steps are required to train a reliable deep learning model. Method In this work, there are two main goals: i) present a fully automated pre-processing algorithm for a smart patch selection within histopathological images, and ii) evaluate the impact of the proposed strategy within a deep learning framework for the detection of prostate and breast cancer. The proposed algorithm is specifically designed to extract patches only on informative regions (i.e., high density of nuclei), most likely representative of where cancer can be detected. Results Our strategy was developed and tested on 1000 hematoxylin and eosin (H&E) stained images of prostate and breast tissue. By combining a stain normalization step and a segmentation-driven patch extraction, the proposed approach is capable of increasing the performance of a computer-aided diagnosis (CAD) system for the detection of prostate cancer (18.61% accuracy improvement) and breast cancer (17.72% accuracy improvement). Conclusion We strongly believe that the integration of the proposed pre-processing steps within deep learning frameworks will allow the achievement of robust and reliable CAD systems. Being based on nuclei detection, this strategy can be easily extended to other glandular tissues (e.g., colon, thyroid, pancreas, etc.) or staining methods (e.g., PAS).