The traditional approach to analysing glass hematoxylin and eosin (H&E) slides is time-consuming and painstaking work. Each slide needs to be examined individually, by a qualified professional. What is more, storage of the slides requires significant physical space, often unavailable within the hospital itself.

So, off the slides go to warehouses, making management and retrieval of specimens difficult and slow. And if a pathologist wants a colleague’s opinion on a specimen? Why then, it will have to be parcelled up in bubble-wrap and couriered to them. It’s almost as though the internet had never been invented.

This would not be such a serious problem if pathologists themselves were not such a scarce resource, especially in the developing world. In some parts of Africa, there is only one pathologist for every 1.5 million people. According to the Chinese Pathologist Association, there are 20,000 licensed pathologists in China, serving a population of more than 1.4 billion.

This shortage of pathologists, coupled with the huge expense of managing glass slide specimens, slows down the process of diagnosis and treatment, with significant impact on patient outcomes.

AI for digital pathology

At Cambridge Consultants, we’re working on ways to address this issue, with AI and hardware projects aimed at digitising and managing pathology specimens in a smarter, swifter way. We believe that the acquisition, management, sharing, and interpretation of pathology information could – and probably should – now be done in a digital manner. What’s needed is high-resolution digital images taken from glass slides, which would be interactive and easy-to-share.

An obvious place to start is oncology, where acquisition and processing of samples is extremely time-sensitive and pathologists are under pressure to provide results to clinicians. One aspect of the process that is a relatively easy win for digitalisation would be to make the initial image-capture systems less expensive.

We are currently developing a low-cost, whole-slide imaging system using more modern, novel imaging modalities that utilise computational optics to overcome some of the existing hardware barriers, which will give multiple readouts. This could offer serious benefits, especially in the places where pathology is often prohibitively costly or simply not available.

Indeed, just think of the possibilities. Once a glass slide image is captured digitally, it could be uploaded to the cloud. A pathology service could then be provided remotely online, with results sent back to the local or regional hospital.

But the issue of pathologist staff hours remains a significant bottleneck. Whether digital or on slides, someone has to analyse the specimens. So far, the impact of AI on pathology has been limited but our view is that AI and machine learning have huge potential.

Once you have digitised images, AI could analyse hundreds of data sets in minutes. Add in a score of the assessment, and the pathologist could only review specimens that fall below a threshold of confidence, saving hours of time. A pathologist with good AI at his or her fingertips is a powerful proposition.

In our work on this area, we have developed a project called BacillAi. The project combines pathology with AI to successfully measure TB infection in patients and monitor progression of the disease. The proof of principle showed great promise and requires an external partner to take this on to meet its full potential.

Having proved that it is possible to measure for a marker or disease, we are now keen to see this process integrated into workflow. Once pathologists and clinicians interact with this information, it could significantly improve patient outcomes.

Computational pathology

Interest in computational pathology is on the increase, but one of the key blockers is the poor generalisability of the current algorithms we have to work with. Several studies in this field have found that algorithms trained on one set of pathology data do not perform well when faced with unfamiliar data.

It is likely this problem is due to the lack of labelled data that AI needs to create robust, generalised algorithms. The problem is that labelled data is expensive and time-consuming to generate, creating a significant problem when developing better algorithms.

We have recently been investigating some alternative approaches. One is to work with semi-supervised learning. In this, AI trains on both small, labelled datasets and also much larger, unlabelled datasets, at the same time. This method leverages the knowledge from the labelled data to extract information from unlabelled data.

One current project we are working on focuses on breast cancer samples from eight classes, four benign and four malignant. This has allowed us to train two models, each with different levels of granularity. One model learns to classify samples as simply malignant or benign, the other learns to classify to the level of each of the eight classes. The results are extremely encouraging.

I firmly believe that now is the time for AI, machine learning and better hardware to revolutionise pathology, saving countless lives. The team here is working hard to make this digital transformation a reality. Please do get in touch if you would be interested in collaborating with us.