Understanding the Potential and Limits of AI in Cancer Care

James Zou, PhD

The field of oncology is witnessing a transformative shift with the increasing use of machine learning and artificial intelligence (AI) being employed at various stages of cancer diagnosis, treatment, and follow-up. According to James Zou, PhD, the use of AI holds potential to revolutionize the way oncologists manage their patients with cancer.

AI is already being used across the spectrum of cancer care through analyzing histopathology images, mammography scans, and other imaging modalities with impressive accuracy. With these tools, clinicians are able to detect disease earlier and provide care in a timely manner.

However, the encouraging potential of AI in oncology comes with some challenges, including data quality, access to results, interpretability, ethical considerations, and more. It will be important for collaboration between researchers, clinicians, and more to fully understand how to best utilize machine learning in the oncology field and better shape the future of cancer care.

In an interview with Targeted Oncology^TM, Zou, assistant professor of biomedical data science at Stanford University, provided an overview of AI and machine learning applications in oncology, addressing the need for further research and development to address challenges and maximize its potential in this space.

Targeted Oncology: Can you discuss machine learning and how it's being utilized in the oncology field?

Zou: Machine learning and artificial intelligence are seeing a lot of applications in oncology. For example, in diagnosis, often the clinicians are working with different kinds of imaging data could be mammography images or CT scans. Machine learning AI algorithms can be very helpful in helping clinicians to analyze those kinds of images for them to identify or to segment relevant regions.

What are some specific examples of how machine learning has been applied into clinical oncology?

There are different stages where machine learning is being applied. They will go all the way from early stages in diagnosis to later stages in terms of treatment planning and treatment recommendations. [On the] diagnosis side, we are seeing a lot of these computer vision algorithms, which is a type of AI or machine learning models that are trained to really understand and analyze different images. For example, now there are algorithms that are looking at histopathology images and slides, and then try to diagnose and predict patient outcomes based on those histology images.

There are also algorithms that are trained to look at mammography images and try to detect tumors, legions from these mammography images as other diagnosis sites and other treatment planning sites. People also develop machine learning models that look at, for example, mutation profiles of patients, right from their somatic mutations, and then try to predict based on these mutation profiles if immunotherapy or some other treatments are likely to be a good treatment for this particular patient.

Could you elaborate on some of the challenges associated with machine learning techniques?

I think there are several challenges. One is that often the models require a reasonably large amount of data to train and to validate. For example, let's say if we're trying to train and develop a machine learning model to analyze histopathology images to diagnose different cancers or diseases, then the model typically requires 1000s, sometimes 10s of 1000s of images that are annotated by clinicians in order to train. It's also possible that these kinds of machine models can pick up and learn about these artifacts or sort of spurious correlations in the images. Maybe it learns about a particular way that images are stained, and it thinks that those kinds of staining techniques are associated with patient outcomes, or it was cancer diagnosis, when those are just sort of technical artifacts. We develop these models, we spend a lot of effort trying to make sure that these models are robust, and they are trying to mitigate these correlations to make sure that they're actually generalizable across different hospitals [and] different conditions.

What are some ways you're addressing these challenges?

One way that we've worked on addressing these challenges first [is] on the data side. We try to curate datasets from quite diverse, heterogeneous sources. To give a concrete example of this, 1 project we've done recently is where we've been creating a lot of these histology, pathology images from Twitter. A lot of the pathologists actually post challenging or ambiguous images that they encounter on social networks like Twitter, so that their colleagues can discuss them. These are our dialogues and conversations. We have actually cured hundreds of 1000s of these pathology images right on from Twitter [and] almost [all] the corresponding discussions by pathologists are on those images. Then we use that data to sort of refine and to train machine learning models to make them more robust, because the models are actually seeing more heterogeneous and diverse kinds of data.

In the review, you came across some machine learning models that have been approved for cancer-related patient usage by regulatory agencies. Could you further discuss this?

I think this is a really interesting [and] exciting space because several of these algorithms and machine learning models have now been approved by the FDA. Just in clinical settings, overall, there's already over 500 FDA-approved medical AI algorithms or medical AI devices. That number is growing quite rapidly. Most of those approvals happened over the last 2 years. Among those 500 FDA-approved AI devices, that includes several algorithms specific for looking at cancer. This includes the algorithms that we discussed for looking at different diagnosis or prognosis are related to cancer patients.

What key features led to their approvals? How have they impacted clinical practice?

In order to get the regulatory approvals for his AI algorithms, the developers will have to demonstrate the algorithms that are highly effective, at least in some combination of retrospective analysis. Sometimes also [they] conduct more prospective studies, although that's not as common as retrospective analysis. By that, I mean that you usually have to show that the algorithm is effective when it's applied to hold-out data, test data from a separate set of patients so that they can make accurate assessments. In addition, I think there's also a lot of interesting discussions now around how to monitor the performance of these AI models, after being approved and then after they've been started to be deployed in hospitals. The key part of that and why that's an important question is because the AI algorithms learn from data. It's a part of the nature of AI. So these are learning algorithms, and the models can improve, or they can even get worse as they learn from more data. Thinking about how to monitor in real time the behaviors and performance of these AI models is another interesting and important topic that we're actively working on.

What potential improvements do you foresee for enhancing the usefulness of machine learning in oncology?

I think we're still at a relatively early stage of machine learning algorithms, AI models being deployed in oncology. I think their potential is very large. Then there's a lot of very rich data, especially oncology. We have data about the text, right from the [electronic health record] of the patients, you have often lots of very rich imaging data and quality. We talked about histopathology, which is often collected in tumor samples, but you also have different X-ray images, CT scans, MRIs, assays, a lot of images, a lot of information. We also have genomics and other lab values like genomic profiling of the tumor samples in these patients.

I think machine learning models can be quite adept at combining these different types of different diverse sources of data in all different stages, all the way from diagnosis to prognosis all the way to treatment, planning, and treatment, design and follow-ups. Across this pipeline, I think there is great potential for machine learning algorithms to come in and make a clinical impact.

What are the key takeaways from this research?

[I hope] our review paper provides an accessible summary of the landscape of AI and machine learning for oncology. It's designed to be read by oncologists or community colleagues so that they can get a quick overview of what is the state of the arts in terms of AI, and maybe some of the AI algorithms could also be useful in their day to day works.

REFERENCE:

Swanson K, Wu E, Zhang A, Alizadeh AA, Zou J. From patterns to patients: Advances in clinical machine learning for cancer diagnosis, prognosis, and treatment. Cell. 2023;186(8):1772-1791. doi:10.1016/j.cell.2023.01.035