Artificial intelligence: a new player in pathology?

by Sambhavi Sneha Kumar

4 January 2022

The field of artificial intelligence (AI) is rapidly advancing, playing an ever increasing role in many disciplines of medicine. Radiology is one such field that was quick to embrace AI. Its use in the analysis of mammograms began as early as the 1990s, with limitations of the human visual system and false negative reports due to oversight being cited as key reasons for welcoming the change. Other fields have been slower to adopt AI, but new advances and applications of technology are paving the way for its increased use. For pathology in particular, pattern recognition methods based on deep learning present a highly promising way to advance and optimise diagnostics and predictions.

Pathology is by no means a stranger to change. Practitioners’ understanding as to what causes disease has been influenced hugely throughout history, as different methods and tools have been developed. Back in the 15th Century, Antonio Benivieni famously recorded case histories and performed autopsies on patients, initiating a progressive wave of endeavours to understand the internal basis of disease. Moving forward through the years, the advent of the microscope became the next driving force in pathology, establishing a cellular basis of disease and increasing the specificity of our understanding of pathogenesis. More recently, our appreciation for the genetic basis of disease and advances in genomics has allowed us to not only characterise certain conditions in the present, but predict future patterns of inheritance. Could increased reliance on AI be the next big shift in the working practises of pathologists?

 

The glory of medicine is that it is constantly moving forward, that there is always more to learn-  William J Mayo

 

In the last decade, there has been a huge surge in research investigating the potential to apply artificial intelligence in image analysis within pathology. Much of this work involved the distinguishing of normal from cancerous tissue in a range of organs including the breast, prostate and cervix, mainly based on processed H&E stained slides. A large proportion of this work involves the use of a particular deep learning algorithm known as a convolutional neural network (CNN). CNNs operate by taking in an input image and assigning features of that image relative ‘importance’, which can subsequently be differentiated. Hence, the network is able to apply multiple filters to extract different features, with subsequent layers of the network focussing on features of increasing complexity as the system learns.

So, what are the advantages of implementing such systems in a clinical environment? Firstly, AI may give rise to improved feature identification. For example, this study compared deep learning based algorithms with expert human pathologists using the Gleason score to analyse histopathological images of prostate cancer specimens. The Gleason score is an established grading system used to evaluate the prognosis of patients with prostate cancer via analysing cell patterns and morphology in biopsies. Interestingly, deep learning systems were able to not only identify features previously identified by expert human pathologists, but also flagged areas of stroma dubbed non-cancerous, suggesting that deep learning systems may be able to spot unique aspects of an image that could be missed by a human. Furthermore, AI systems can ‘learn’ rapidly from a huge bank of images, continuously refining their diagnostic criteria, whereas it would take a human far longer to accrue the same magnitude of experience. AI systems also circumvent other issues associated with human working practises, such as fatigue or bias. The future of pathology may well involve humans working alongside computers. AI systems can perform the repetitive but meticulous tasks for which humans may be error prone (e.g. counting mitotic figures in a cell) whilst pathologists focus on tasks that may be more cognitively demanding. This idea of AI working with humans to increase overall efficiency is starting to be trialled at the clinical level. 

Whilst it is true that pathology is no stranger to change, any major development of working practises will pose challenges. In order to optimise the training for deep neural networks, a large quantity of data is needed. For rarer diagnoses and unusual circumstances there will likely be a shortage of the relevant images. AI systems, arguably, are not actually as ‘intelligent’ as humans in some ways; they work within limited constraints of rules without independent thought. Hence, their ability to analyse a scenario for which they have not been trained is virtually non-existent. Furthermore, the input needed for a clinical diagnosis extends beyond image analysis for tissue samples or scans. The ‘perfect’ AI system in pathology may have to integrate data from a large range of sources including images, patient history and any genetic data available. 

Furthermore, AI systems, which are still arguably being refined, will need to be rapidly integrated into training programmes, and furthermore, existing pathologists may need to be given access to training. Deep learning systems will undoubtedly make errors, especially towards the beginning of their implementation, and refinement will depend upon communication between pathologists and AI time. This will limit productivity and possibly cause new friction in the working environment. Furthermore, there are considerations regarding the issue of culpability: if a diagnostic algorithm makes a mistake that affects a patient’s wellbeing, who should be held accountable? This question has already been debated in many other contexts (such as driverless cars), and can be difficult to answer.

The role of a pathologist is not purely diagnostic; taking samples, managing laboratories, contributing a clinical perspective and dialogue with patients are also parts of the role. Artificial intelligence undoubtedly represents an important and efficient tool for pathology, and will almost certainly go on to play a key role within the field in the future – but this shift will likely be gradual and will take some time. It remains to be seen how widely implemented these systems will become – both within the U.K. and from a global perspective – and how successful they will be in clinical practice.

 

Image credit: kkolosov, pixabay.com