Computer-aided Diagnosis: The Tipping Point for Digital Pathology

For some 10 years now, digital pathology’s proponents within the industry have regarded it as the “next big thing.” Yet today, there are still scarcely more than a handful of pathology laboratories fully equipped to digitize their workflow. The reasons for this scarcity are mostly legal or financial in nature. Especially in the US, the “struggle” with the FDA over regulation prevents large-scale introduction of whole slide images (WSI) into pathology. Nevertheless, it remains to be seen whether many pathology labs would invest the required (large) amount of money in a technique that still has to substantiate its high expectations.

Will WSI be the disruptive innovation, as some opinion leaders state? I think it will, but only under the right conditions. Yes, it may streamline the laboratories’ diagnostic workflow, but only if it is fully integrated within the larger pathology department as a whole. System efficiency may be increased only by proper integration with current workflows and information systems. WSI may be disruptive in the way we practice pathology at large, but only if we are able to create the right infrastructure to support networks of collaborating pathologists. Then, it will be possible to practice “pathology in the cloud” and instantly reach the right (subspecialized) pathologist for every difficult case.

Disruptive as this may be from an organizational perspective, perhaps the most important promise of digital pathology is the development of computer-aided diagnosis (CAD) algorithms for WSI. Once we are capable of having a computer help a pathologist in interpretation of histopathological images, digital pathology will realize its full potential. Almost all research so far has been devoted to quantifying immunohistochemistry. The next step will be assessment of regular H&E stained sections.

I believe that right now, there are three (overlapping) areas in which CAD will have a major impact on pathology diagnostics, with increasing levels of impact and complexity.

  1. CAD will increase efficiency of routine tasks. The first use cases for CAD in pathology will most likely address tedious routine diagnostic tasks that require great accuracy, such as finding metastases in lymph node sections. Most pathologists are not fond of this part of their work, yet it needs to be done well as it is of high relevance for tumor staging. A CAD algorithm that can detect metastases fully automatically will alleviate this task while actually increasing accuracy.
  2. CAD can improve accuracy of tasks in which some grading is involved. It is widely known that pathologists possess – at best – moderate reproducibility in such (semi-) quantitative tasks. A well-known example is Gleason scoring for prostate cancer. CAD may offer a powerful alternative, quantifying tissue changes that correspond with tumor grade in an accurate and reproducible manner.
  3. CAD may yield relevant information, for diagnosis and prognosis, which the human eye and mind are unable to recognize or appreciate. Instead of using a computer to mimic a pathologist in, for instance, grading a tumor, we could also try to obtain relevant quantitative data directly from WSI. These so-called “imaging biomarkers” may drastically change the way we extract information from tissue sections. While promising, it will take a significant amount of research and validation before the patient benefits from this type of application.

The research community will have to develop and validate algorithms before CAD is really going to take off. Deep learning, a modern pattern recognition technique, has been shown to be extremely powerful in many different disciplines and for a wide variety of problems. We have recently shown (Litjens et al. Scientific Reports 2016) that deep learning is also particularly suited for analyzing WSI. We found that deep learning yielded CAD systems that are close to being clinically useful.

To get a broader view on the field and establish the current state-of-the-art of CAD in Pathology for a specific application, we organized the “Camelyon16” grand challenge (http://camelyon16.grand-challenge.org). A grand challenge is a valuable instrument in the field of medical image analysis, in which every researcher or research group is invited to develop an algorithm for a given problem. All participants solve the same problem using the same set of data, and the challenge organizers evaluate all submissions in exactly the same manner. It is, therefore, a great way to compare different approaches to solving a problem.

In the Camelyon16 challenge, we offered the participants a large number of full WSI depicting sentinel lymph node sections of breast cancer patients with exhaustive annotation of all metastases. We collected tissue sections in two different Dutch hospitals and scanned these on two different WSI scanners. Participants used these WSI to develop CAD algorithms. In the next stage of the challenge, participants ran their algorithms on a separate set of 130 sentinel lymph node WSI (this time without the “ground truth” annotations) and results were sent to us for evaluation. An important result of our challenge is that many strong research groups and companies got to work on a very specific, clinically relevant application in histopathology, with a large set of fully annotated WSI, in a direct comparison. This has definitely increased the rate of developments in this field, challenging researchers to spend significant amounts of time on a problem they would otherwise not tackle.

The result is a number of strong algorithms for this task. Deep learning was applied in all of the top 10 algorithms in this challenge, underlining the superiority of this technique for CAD in pathology. We were excited to see that the top algorithms in Camelyon16 performed at the same level of accuracy as a pathologist participating in the challenge. This was somewhat to our surprise: we expected CAD to be a powerful and promising development but we hadn’t expected this level of performance at this stage. While promising, before we can start implementing CAD we have to validate these algorithms rigorously on large numbers of cases.

It is my strong conviction that implementing CAD will tip the business case completely, and may even improve pathology diagnostics. Before we reach that point, a significant amount of work has to be carried out. Development and validation of CAD algorithms requires, more than anything, involvement of expert pathologists. The rewards, however, are huge: a modern, accurate and reproducible assessment of human tissues facilitating the best treatment of every individual patient.

Dr Jeroen van der Laak, PhD (Jeroen.vanderlaak@radboudumc.nl) is associate professor in Digital Pathology at the Radboud University Medical Center in Nijmegen, The Netherlands. He has a background in computer science and leads a research group on computational Pathology. Dr van der Laak is the initiator of the CAMELYON16 grand challenge, which is the first challenge on CAD in pathology using whole slide images. At the time of writing, his team is organizing the successor CAMELYON17. Dr van der Laak is board member of DPA, organizer of a computational Pathology session at the European Congress of Pathology and is frequently invited as a (keynote) speaker for conferences in (digital) pathology.

In seeking to foster discourse on a wide array of ideas, the Digital Pathology Association believes that it is important to share a range of prominent industry viewpoints. This article does not necessarily express the viewpoints of the DPA; however, we view this as a valuable point with which to facilitate discussion.

Leave a Reply

Your email address will not be published. Required fields are marked *

18 + 19 =