How to implement machine learning to reap true advantages in pathology
Posted 16th August 2017 by Jane Williams
Digitization paves the way for machine learning
With the first FDA approval for primary diagnosis now being a fact, together with several European labs performing the majority of their review digitally, the digitization of pathology has sped up and is now in many labs’ plans.
Digital pathology brings benefits like the sharing images to colleagues within seconds and more efficient preparing and presenting at tumour boards. But the biggest impact in the long term is probably the enabling of applying Machine Learning (ML) based image analysis (including deep learning). These tools will not only speed up the diagnosis itself but also increase the precision, leading to a selection of the right treatment patients. These benefits that come with ML tools will justify the business case for many labs to go digital.
But for machine learning applications to really impact pathologists’ daily work they need to be implemented with workflow in mind. In this article, I give a brief explanation of ML in general, its use for pathologists and advice on what to consider when implementing ML in your pathology operations.
What is machine learning?
To put it simply, ML solutions automatically create and learns a programme from data. Since the data is teaching the programme, we do not need to define each possible step or need to know how the programme works (e.g. if this happens, then do that).
Instead, you supply the algorithm with a set of examples, for instance in the pathology case a set of images of a specific tumour, or images with annotated features such as mitoses (dividing cells). The programme learns from these images, and when given a new image it can classify the tumour type or identify mitosis based on the example set.
We all benefit from machine learning already today
ML is already part of most peoples’ lives. For example, when writing this text, I am listening to my “Discover weekly” playlist on Spotify, which is created by an ML programme. The algorithm is given example songs from my personal created playlists as well as previous songs I have listened to. Based on those examples it creates a list of songs that Spotify thinks I am likely to enjoy.
As we all have experienced by now, these ML programmes are far from perfect. They make mistakes. This is highly important to take into consideration when applying ML in pathology. There needs to be a plan on how to detect and handle the errors made by ML applications.
How can machine learning help the pathologist?
In the long term, ML will revolutionise cancer care. In the short term it will probably not take long before ML will impact and facilitate the daily tasks of a pathologist. A few examples:
- In a similar way as Google rank results when you use its search engine, a digital pathology system could create a ranking of likely answers on a specific question. For example, what tissue type or disease a marked area is.
- As Instagram suggest persons that you might like to follow, the system could give recommendations to a specific problem. For example, “Have you tried to use a Giemsa staining to look for helicobacters…” (bad example, but you get it).
- Clustering is also a powerful approach that is used when you shop online and the web shop gives examples on “other persons also looked at…”. This function would provide the pathologist with insights on what colleagues have found out from similar cases, like a virtual second opinion.
- Notify the pathologist about uncommon features in the image can be compared to how Twitter or Facebook can identify “trending” topics and products and inform you about these.
The glorification of ML & its limitations
So will ML replace the pathologist? The answer is probably no. The new technology will most likely only replace a small component in a pathologist’s daily work. But those components that ML will replace will be crucial parts in a patient’s care where the computer has a significant advantage over the human. For example, in the task of counting cells responding positive to a specific staining. Efficient and precise quantifications will become even more important as immunotherapy moves forward where the pathologist will play a key role in determining which treatment that will be given to the patient.
It is also crucial to lower the hurdle to apply ML technology so the effort is significantly below the perceived value of the benefits. One example is the voice recognition used by Google and Apple. 2-3 years ago they worked partly well, but still required some manual work which created a lot of frustration, leading to few users. Today voice recognition has reached a level where they save time for the user, hence the number of users have increased significantly.
Not the last is the legal aspect of using ML. As long as the ML only give suggestions and advice and not decide on the diagnosis and the treatment the regulatory hurdle is much lower. By applying ML based on suggestions will allow for error handling and ease its roll-out to healthcare so patients and pathologists can utilise the new technology faster. This philosophy is for example in line with that of IBM’s Dr Watson.
Key advice when implementing machine learning in pathology
To keep an efficient workflow is key when using ML applications in a digital pathology. The algorithm should work “behind the scenes” and there is little patience for waiting for processing or changing application. Therefore, a tight integration with the viewer is absolutely crucial to lower the hurdle of use and increase efficiency. A key enabler for that integration is an open, vendor neutral digital pathology solution. The digital pathology solution you choose should have a good ML support for the most commonly performed tasks, but in addition, the solution must have an open, standards based interface so that you can implement the ML applications you want regardless of vendor.
A second key aspect is to let the pathologist still be in charge of the decision and give him or her the ability to interact with the ML application. As stated before no algorithm is perfect, hence giving the pathologist the opportunity to change the result will create a feeling of command and put the pathologist in the driver seat instead of the opposite. For this to become reality the visualisation of the result is key to illustrate how the algorithm has based its decision. One example is the cell counting below where each positive cell is marked with a + sign, and the pathologist can change those to a – if they judge it is not positive. The ML application should remember corrections to create incentives for the pathologist to interact and refine the tool.
ML’s introduction to pathology – an enormous pie that will be eaten in small chunks
The opportunities with ML are endless and the technology is evolving very fast. In the consumer industry, its use is quickly increasing. In a similar way, ML will facilitate the daily lives of pathologists.
However, due to the conservative nature of healthcare, and the severity if an error would occur, ML algorithms for pathology should focus on the small parts where the computer has an advantage over the human, where the hurdle to use the programme can be kept low, and where its use is easy to adopt from a legal perspective. The ML algorithm should put the pathologist into the driver seat.
See who is speaking at the event. View the agenda here.