Jeffrey J. Nirschl, Andrew Janowczyk, Eliot G. Peyster, Renee Frank, Kenneth B.
Margulies, Michael D. Feldman, Anant Madabhushi,
Chapter 8 - Deep Learning Tissue Segmentation in Cardiac Histopathology Images, Editor(s): S. Kevin Zhou, Hayit Greenspan, Dinggang Shen, Deep Learning for Medical Image Analysis, Academic Press, 2017, Pages 179-195, ISBN 9780128104088, https://doi.org/10.1016/B978-0-12-810408-8.00011-0. (http://www.sciencedirect.com/science/article/pii/B9780128104088000110) Abstract: Cardiac biopsies are obtained for the diagnosis of acute heart failure as well as post-heart-transplant rejection. On histology, both are characterized by an expansion of the cellular and acellular stromal tissue as well as morphologic changes within myocytes, among other features. There is growing evidence, from a number of diseases, that computer-extracted features of tissue and nuclear architecture are not only useful for diagnostics, but can also predict disease recurrence, progression, and patient outcome. Automated tools to segment myocytes and stroma with high fidelity are a critical first step toward developing predictive models to improve diagnosis, prognosis, and reproducibility in cardiac pathology. Whole-slide imaging in digital pathology is producing large, rich data sets for machine learning algorithms that learn features representations directly from data, such as deep neural networks or “deep learning”. Deep learning has achieved state of the art performance in computer vision and is becoming ubiquitous within the realm of digital pathology. Here, we present a deep learning framework for myocyte and stroma segmentation in H&E stained cardiac biopsies using a data set of 103 patients with clinical heart failure or non-heart failure controls. We provide a tutorial for how to use the AlexNet architecture, implemented in Caffe, to train a pixel-level classifier for segmentation. We compare the performance of our deep learning approach to a random forest classifier with 333 intensity and texture features and test on 20 held-out patient images with expert ground-truth annotation of tissue boundaries. We find our deep learning segmentation outperforms random forests in terms of mean (a) AUC (0.95 vs 0.80), (b) F-score (0.96 vs 0.91), (c) TPR (0.96 vs 0.92), (d) TNR (0.94 vs 0.67), and (e) modified Hausdorff distance on the segmentation perimeter pixels (51 vs 122). Keywords: AlexNet; Caffe; Cardiology; Convolutional neural networks; Computer-aided diagnosis (CAD); Deep learning; Digital pathology; Cardiac biopsy; Heart failure; Hematoxylin and eosin (H&E); Whole-slide imaging (WSI)