It’s true that some pathologists fear the loss of their jobs to computer algorithms and other artificial intelligence (AI) technologies. But how reasonable is that fear? How likely is it that diagnostic algorithms will be able to perform the tasks that, up until now, have required the eye and intuition of a human pathologist? And what, ultimately, will be the role of the machine in the pathology laboratory of the future?The first step to implementing AI in the pathology workflow is to deploy digital pathology, which involves scanning glass pathology slides and storing the resulting images in a digital repository to be viewed on computer screens. As soon as a digital image has been acquired, AI algorithms can analyze the data and provide additional information. The pathologist can have the results ready as soon as they start reviewing the images, meaning that they can use that additional information data to provide a deeper, more nuanced diagnosis.
Advances in slide scanning technology and cost reduction in digital storage capacity are now enabling full digitalization of the microscopic evaluation of stained tissue sections. One of the most significant benefits of these technological advancements is the possibility for computer-aided or computerized diagnosis – not for every patient or every disease, but in a way that permits pathologists to focus their attention on the cases that most need it. AI techniques enable computers to solve perceptual problems, such as image recognition, which could lead to drastic improvements in pathology diagnostics in terms of objectivity, accuracy, and efficiency. In my opinion, AI is not a replacement, but an excellent support system for pathologists.
The CAMELYON16 challenge
My colleagues and I organized the CAMELYON16 challenge (CAncer MEtastases in LYmph nOdes challeNge) to evaluate state-of-the-art machine learning methods for the detection of metastases in sentinel lymph node tissue sections – and to compare their performance with trained pathologists. The majority of the algorithms submitted for the challenge were based on deep artificial neural networks, which are at the forefront of machine learning algorithms. Deep neural networks consist of multiple layers of interconnected artificial neurons; these algorithms seek to draw a relationship between an input (for instance, a diagnostic image) to an output (“cancer” or “benign”). To build the system, we expose the deep learning algorithm to a large dataset of labeled images, with which it teaches itself to identify relevant objects. During the learning process, connections in the deep neural network become stronger or weaker as needed to make the system better at prediction. After training, we can apply the network to images it has never “seen” before. CAMELYON16 was the first grand challenge in pathology to offer a large collection of whole slide images (a total of 700 GB of image data!), helping researchers around the world bridge the gap from working on small region-of-interest images to whole slide scans. In this challenge, every participant (whether digital algorithm or human pathologist) was given one single slide per patient from which to determine the presence or absence of breast cancer metastases. In a real clinical setting, however, we would evaluate sections from multiple levels. The test dataset was also enriched with cases containing metastases. The majority of specimens pathologists encounter in daily practice do not contain metastases – so this enrichment was necessary to achieve a well-rounded representation of what might be encountered in clinical practice without including an exorbitant number of slides or several terabytes of total data. At the moment, the majority of pathologists worldwide interpret pathology specimens under a microscope. The reason challenges like CAMELYON16 – and now CAMELYON17 – exist is because the next generation of pathologists will be working with digital images day in and day out. We’ll need appropriate algorithms to help them manage the sheer volume of data they will encounter, and we’ll also need to ensure that these pathologists are trained to use AI and computational methods in their daily practice.A new assistant
The results of our article on the diagnostic assessment of deep learning algorithms for metastasis detection (1) showed that there is a significant opportunity for AI in pathology – namely, assisting pathologists with the interpretation of histopathologic sections. I suspect that, as technologies advance and adoption increases, we will see more applications of AI in pathology in diagnosis, treatment, outcome prediction and prognosis evaluation.
AI is increasingly being recognized as a major element of the overall healthcare landscape. We are now at a turning point where computers perform better than clinicians at specific tasks. This offers a great opportunity to empower clinicians by improving their efficiency and accuracy. That’s not to say that I think AI will replace clinicians; rather, it will gradually change the way they work. Pathologists may, for example, spend less time on the interpretation of pathology slides, and instead focus on more critically important tasks, such as aggregating data from multiple sources (molecular, genetic, radiological, and so on) to better understand patterns that lead to more accurate and definitive personalized diagnosis. Rather that neglecting AI or feeling threatened by it, I think medical professionals – and pathologists, in particular – should embrace AI solutions. There’s a bright future ahead for human-computer collaboration!
Newsletters
Receive the latest pathology news, personalities, education, and career development – weekly to your inbox.
References
- B Ehteshami Bejnordi et al., “Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer”, JAMA, 318, 2199–2210 (2017). PMID: 29234806.
