Cancer Diagnostics in the Age of AI

Whether you love or loathe artificial intelligence (AI), the medical regulatory bodies have certainly recognized its value. The FDA alone has authorized over 800 AI or machine learning-enabled medical devices (1).

And there is no escaping the fact that this recognition now includes diagnostic tools. In September 2021, the FDA approved the first AI-based software for cancer diagnostics. The tool was trained to tell, with a high degree of accuracy, whether prostate cancer was present or not in a digital slide from a prostate biopsy histological preparation. If cancer is present, it reveals the suspicious area – for further review by the pathologist.

As AI tools begin to gain traction for use in cancer diagnostics, we asked a manufacturer and an end user to share their views on the difference AI makes to pathology practice.

We spoke with Juan Retamero, anatomical pathologist and Medical Vice President of Pathology Operations and Diagnostic Products at Paige. Following the 2021 approval, Paige received an FDA breakthrough designation for a breast cancer diagnostic algorithm in 2023. The designation recognizes that the algorithm demonstrates value to clinical practice, and qualifies it for an expedited program for FDA approval.

Please outline the study that led to the FDA breakthrough designation of your breast cancer algorithm…
 

The breakthrough designation was awarded in response to a validation study in which we quantified the performance of pathologists who reviewed a data set of lymph nodes from patients with breast cancer (2). The pathologists reviewed the same data set twice, both with and without AI assistance.

When the pathologists reviewed the data set with AI assistance, their sensitivity increased – they picked up even smaller metastases and the accuracy of their diagnoses improved. We also found that, with AI assistance, they reviewed the data set in half the time – or became twice as efficient.

Why is this big news for patients and pathology labs?
 

The detection of lymph node metastases is essential for breast cancer staging, and missing them can profoundly alter the patient’s stage and thus affect the treatment pathway, prognosis, and long-term outcome. We have shown that AI assistance improves diagnostic accuracy; hence patients benefit.

From the laboratory perspective, our results show that AI can help pathologists do their diagnostic work faster and more reliably in a task like lymph node examination, which is labor intensive and time consuming – and can also be quite arduous when done on a traditional microscope. Carrying out the task with a digital screen and a computer is more ergonomic and a lot more comfortable. 

The AI saves time because it recommends an area of the image to focus on, based on where it detects cancerous cells. This makes the task much more focused, faster, and more accurate, as seen in our study.

How does AI support cancer grading investigations?
 

Immunohistochemistry (IHC) tests are great for identifying cancer biomarkers in biopsy samples, but treatment is only offered to patients whose biomarker-positive cells exceed certain thresholds. Pathologists usually perform this quantification by eye, estimating the number of positive cells. For those patients with results very close to the treatment threshold, this means their management is determined from a subjective test.

Today, a computer can quantify the biomarker cells more accurately than a human, and tell you which side of the threshold the patient happens to be. The results are much less subjective, and much faster. In other words, patients receive better classifications of their biomarker status, leading to better informed treatment decisions by the oncologist.

What difference do you think this technology could make to diagnostic turnaround times?
 

Based on the experience from our studies, using data from real users in labs, we see that pathologists’ efficiency increases by 25 to 50 percent when they are aided by these tools. In other words, they take between 25 and 50 percent less time to do the same tasks when they are aided by AI.

What would you say to alleviate the fears of AI skeptics?
 

I was not practicing pathology in the 1980s, when IHC emerged, but I’m aware that its advent was received with some reluctance and skepticism. The truth in pathology lay on the H&E slides that pathologists had known and trusted for decades.

But eventually people did realize that, in some situations, IHC was quite helpful. For example, in the past it might have been difficult to affiliate a metastasis from an unknown primary to, say, the lymph nodes. Now, immunostains allow us to say, with reasonable confidence, which organ the metastasis originates from. In time, people became aware that IHC was not intended to replace current practice, but rather to enhance it – to increase the efficiency and accuracy of the diagnosis.

The same thing is happening now with molecular pathology. Pathologists might maintain that there will always be classic surgical pathology problems for which molecular assays are not going to be helpful. However, labs increasingly recognize that molecular diagnostics offer another tool with which to classify disease in order to provide the patients with the best available treatment.

AI is once again raising the same sorts of concerns. Pathologists might fear that AI is here to radically change current practice and to render them obsolete. But this is not the case. It’s only another adjunct or additional tool to keep in your armory to help you in certain situations. It’s not a silver bullet and it’s certainly not here to replace pathologists, but rather to help pathologists work more efficiently and more accurately.

How might AI address the issue of pathologist burnout?
 

I think the profession should embrace any help that it can get, because what we do is difficult and there aren’t enough of us around. Pathologists are retiring at a faster rate than we are training junior pathologists. At the same time, the number of cancer cases is increasing, in line with the aging population. Additionally, as screening tests improve and more investment is made in screening programs around the world, laboratory workloads are increasing. AI can help to absorb some of this additional burden by facilitating some tasks and increasing workflow efficiencies.

I like to view AI as a digital trainee, much like a fellow, a registrar, or a resident pathologist. It previews cancer cases on behalf of the pathologist, prioritizes those cases that are likely to be malignant, and moves those to the top of the list for review by the pathologist. Multimodal AI tools, that can do both image analysis and text generation, also prepare a draft of the laboratory report. As the end result of the pathology diagnostic process, the lab report often takes pathologists a significant amount of time to prepare and validate. Then, when the pathologist picks up the previewed case, their job is to verify what the AI has done. If everything looks correct, it can simply be validated and passed on to the clinician.

Like any other diagnostic tool, AI is not perfect. It might generate some false positives, false negatives, and inaccuracies. And that is why the pathologist will always be required – to make sense of what the AI is telling us, correct any inaccuracies, and provide a report that makes sense.

This is where the synchronicity between humans and AI becomes very interesting; pathologists using this technology perform better than pathologists who are not using it, as shown in many of the peer-reviewed publications generated by pathologists using this technology.