Journal of Pathology Informatics Journal of Pathology Informatics
Contact us | Home | Login   |  Users Online: 307  Print this pageEmail this pageSmall font sizeDefault font sizeIncrease font size 

Table of Contents    
J Pathol Inform 2018,  9:25

Deep learning for medical image analysis

Center for Biomedical Informatics, Wake Forest School of Medicine, Winston-Salem, NC 27157, USA

Date of Submission25-Apr-2018
Date of Acceptance22-May-2018
Date of Web Publication25-Jun-2018

Correspondence Address:
Dr. Metin Nafi Gurcan
Center for Biomedical Informatics, Wake Forest School of Medicine, Winston-salem NC 27157
Login to access the Email id

Source of Support: None, Conflict of Interest: None

DOI: 10.4103/jpi.jpi_27_18

Rights and Permissions

How to cite this article:
Senaras C, Gurcan MN. Deep learning for medical image analysis. J Pathol Inform 2018;9:25

How to cite this URL:
Senaras C, Gurcan MN. Deep learning for medical image analysis. J Pathol Inform [serial online] 2018 [cited 2018 Jul 17];9:25. Available from:

Edited by: S. Kevin Zhou, Hayit Greenspan, Dinggang ShenAcademic press 2017

Pages: 433 pages

ISBN: 978-0-12-810408-8

   Introduction Top

Recent successes in deep learning have ushered in a new level of excitement on what computers can do in our lives. Deep learning approaches have not only brought many important innovations such as self-driving cars, digital assistants, and financial forecasting but also revolutionized approaches to image analysis and computer vision. Nowadays, several research groups are developing deep learning approaches for many different medical image analysis problems. However, deep learning approaches have not been properly introduced to a large segment of the research community – several misconceptions are common about what deep learning is and what it can (and can't) do. Recently, a reviewer of our proposal, which explored the use of neural networks for bladder histopathology analysis, warned us that “there are no neurons in the bladder!” Luckily, there are some books to inform researchers about this rapidly developing and exciting field.

Deep Learning for Medical Image Analysis, edited by Zhou, Greenspan, and Shen, is a recently published book providing background on deep learning and its application to several medical image analysis problems. The book is not a comprehensive reference book; instead, the editors carefully selected different novel studies to demonstrate advantages of deep learning and to inspire readers to apply deep learning methods to their image analysis problems.

The book is divided into six parts; the first part gives an introduction to neural networks and convolutional neural networks (CNN), a popular form of deep learning for image analysis. This introduction is relatively high level and assumes that the reader has an excellent mathematical background and some experience in image analysis. A brief, introductory-level description of the neural networks would have been useful for the readers new to the topic before describing mathematical details.

The second part includes five interesting chapters that describe how to use deep learning for image detection and recognition tasks. In addition to two chapters focusing on mitosis detection and nucleus localization for digital histopathology, CNN is applied to many different radiologic imaging techniques such as magnetic resonance (MR) imaging, computerized tomography, and video ultrasonography.

The third part focuses on segmentation problems and includes three chapters that describe the application of deep learning approaches to histopathology and MR images. Interestingly, one of the chapters compares segmentation results of deep learning features with those of classical hand-crafted features. Although the conclusions of this chapter cannot be generalized, they would be interesting for those who feel more at home with traditional image analysis techniques. While these chapters provide a quite comprehensive look at applications of neural networks, the coverage would have been improved by including some recent fully convolutional network-based studies. Because these types of networks are more suited to semantic segmentation, the algorithms may work faster than traditional CNN solutions.

Image registration is another challenging topic for medical image analysis, and the fourth part includes two relevant examples which show how deep learning can be applied to this problem. In the first example, a novel approach is presented by combining state-of-the-art image registration solutions with deep learning feature representations. The other example demonstrates how CNN can register two- and three-dimensional images in real time. Given the success of these applications, it is likely that we will see many deep learning-based image registration approaches in the near future.

The chapters in the last two sections demonstrate many diverse applications of deep learning for biomedical images: using two unregistered images to classify mammograms; applying transfer learning for chest radiology-pathology categorization; and showing the predictive power in identifying future disease progression for Alzheimer's disease.

Although many image analysis books in the past focused entirely on radiological imaging, it is exciting to see that this book contains histopathology applications alongside radiological applications. While some chapters treat both fields, there are chapters (e.g., Chapters 7 and 8) that mainly focus on microscopic imaging using both H&E and immunohistochemistry staining. The last chapter is dedicated to natural language processing: specifically, to process radiology reports to annotate and analyze radiological images with minimum human efforts. Although the application is geared toward radiological imaging, similar techniques could also be applied to pathological imaging and analysis.

Three of the most frequently asked questions by those who are new to deep learning are (1) how to select the correct framework (i.e., which network is better and how to select the parameters); (2) whether deep learning approaches are like black boxes; and (3) whether deep learning approaches will replace pathologists. These questions are not explicitly discussed in the book. To address the first question, it should be noted that, in general, researchers empirically design networks or employ popular networks and then apply transfer learning. New exciting smart systems are emerging, and these systems are able to design themselves. This excitement should be moderated with the fact that these systems are not magical: their performance highly depends on the quality and quantity of the training samples. Improper dataset design or improper ground truth may result in poor performance or generalizability issues. As for the second question, researchers often think that deep learning networks are black boxes and that nobody knows what is going on in these artificial neurons. Actually, there any many ways of visualizing deep learning networks, and these methods may help researchers better understand how a trained network functions. Some researchers firmly believe that deep learning will become so powerful that we will not need pathologists within the next decade. While this discussion evolves, it is safe to assume that deep learning will be very beneficial to pathologists for decision support in the near future.

In summary, this book is a good – though slightly high level – general introduction for medical image analysts interested in deep learning. The chapters were carefully selected by the editor to show the variability of deep learning approaches and their applicability to various image analysis problems.




   Browse articles
    Similar in PUBMED
   Search Pubmed for
   Search in Google Scholar for
    Access Statistics
    Email Alert *
    Add to My List *
* Registration required (free)  

  In this article

 Article Access Statistics
    PDF Downloaded164    
    Comments [Add]    

Recommend this journal