Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases.

摘要: Background: Deep learning (DL) is a representation approach ideally suited for image analysis challenges in digital pathology (DP). The variety of tasks the context DP includes detection and counting (e.g., mitotic events), segmentation nuclei), tissue classification cancerous vs. non-cancerous). Unfortunately, issues with slide preparation, variations staining scanning across sites, vendor platforms, as well biological variance, such presentation different grades disease, make these particularly challenging. Traditional approaches, wherein domain-specific cues are manually identified developed into task-specific handcrafted features, can require extensive tuning to accommodate variances. However, DL takes more domain agnostic combining both feature discovery implementation maximally discriminate between classes interest. While approaches have performed few related tasks, classification, currently available open source tools tutorials do not provide guidance on (a) selecting appropriate magnification, (b) managing errors annotations training (or learning) dataset, (c) identifying suitable set containing information rich exemplars. These foundational concepts, which needed successfully translate paradigm non-trivial (i) experts minimal histology experience, (ii) processing derive their own, thus meriting dedicated tutorial. Aims: This paper investigates concepts through seven unique use cases elucidate techniques produce comparable, many cases, superior results from state-of-the-art hand-crafted feature-based approaches. Results : Specifically, this tutorial analysis, we show how an framework (Caffe), singular network architecture, be used address: nuclei ( F -score 0.83 12,000 epithelium 0.84 1735 regions), tubule 795 tubules), (d) lymphocyte 0.90 3064 lymphocytes), (e) mitosis 0.53 550 (f) invasive ductal carcinoma 0.7648 50 k testing patches), (g) lymphoma (classification accuracy 0.97 374 images). Conclusion: represents largest comprehensive study date, over 1200 images during evaluation. supplemental online material that accompanies consists step-by-step instructions usage supplied code, trained models, input data.