Ruijiang Li, PhD
The promise of artificial intelligence (AI) technologies to provide highly personalized oncology care for patients and improve outcomes has been decades in the making. In a 1987 editorial in The New England Journal of Medicine, pioneering nephrologist and health economist William B. Schwartz, MD, and his colleagues wrote about the expectations of AI use in medicine and wondered why they hadn’t materialized yet.
“After hearing for several decades that computers will soon be able to assist with difficult diagnoses, the practicing physician may well wonder why the revolution has not occurred. Skepticism at this point is understandable. Few, if any, programs currently have active roles as consultants to physicians. The story behind these unfulfilled expectations is instructive and, we believe, offers hope for the future.”1
Although the expectation of AI to completely revolutionize cancer care has yet to be fully realized, its use in oncology is no longer hypothetical. The pace of advancements in AI tools in clinical practice has quickened, especially in the areas of cancer diagnostics; computer vision, to automate tasks that replicate human capabilities; risk assessment; patient prognosis estimation; and treatment selection. Progress in methods and algorithms for training AI models, the development of computer hardware needed to train these models, and access to large volumes of cancer data—including imaging, genomics, and clinical data—have all converged to produce new applications for AI in cancer research and clinical care.
These advances are already starting to produce tangible results. For example, researchers at the National Cancer Institute (NCI) have developed an AI model that uses data from individual tumor cells to predict drug responses using widely available bulk RNA sequencing data, and then they fine-tuned that model using single-cell RNA sequencing data. Called the Personalized Single-Cell Expression-based Planning for Treatments in Oncology (PERCEPTION), a precision oncology computational pipeline, the researchers used this approach to build AI models for 44 United States Food and Drug Administration (FDA)–approved cancer drugs. In a proof-of-concept study, they found that the models accurately predicted how individual cells would respond to both single drugs and combinations of drugs. PERCEPTION also demonstrated accuracy in predicting responses to combinations of drugs in clinical trials for multiple myeloma and breast cancer and was successful in capturing the development of resistance in patients with lung cancer who were treated with tyrosine kinase inhibitors.2
Predicting Treatment Response
Achieving truly personalized care for individual patients with cancer is also a major research focus of Ruijiang Li, PhD, Associate Professor of Radiation Oncology, Department of Radiation Physics, Stanford University, and member of Stanford Bio-X, the University’s pioneering interdisciplinary biosciences institute. In his study presented during the 2024 ASCO Annual Meeting investigating the use of AI-based single-cell analysis of digitized whole-slide hematoxylin and eosin (H&E) images to predict objective response and survival benefit from anti–PD-1/PD-L1 immune checkpoint inhibitors in patients with advanced gastroesophageal cancer, the results show the approach was successful in predicting treatment outcomes in these patients.3
In the study, Dr. Li and his colleagues obtained whole-slide H&E images from 82 patients with advanced gastroesophageal cancer that had been treated with anti–PD-1/PD-L1 immune checkpoint inhibitors at Stanford University in California for discovery, and 189 patients with advanced gastric adenocarcinoma treated at Southern Medical University in China were used as an external validation cohort.
The researchers developed a fully automated cell annotation approach by leveraging multiplex immunofluorescence and then trained a deep learning model to classify nuclei into four cell types from H&E slides: tumor cells, lymphocytes, neutrophils, and macrophages. A total of 66 features were computed to quantify cell composition and cell-cell interactions within the tumor microenvironment.
“Immune checkpoint inhibitors are the most exciting development in cancer treatment in the past decade, and they have become the standard of care for patients, not just in the treatment of gastrointestinal cancers, but in most solid tumor cancers as well,” said Dr. Li. “What we were able to show in our study with this AI-based single-cell approach is that for predicting objective response in the validation cohort, a multivariate model combining the spatial features achieved an area under the curve [AUC] of 0.81 compared with an [AUC] of 0.65 for PD-L1 combined positive score (CPS) [P = .0014]. Combining the multivariate model with PD-L1 CPS achieved an [AUC] of 0.84, and that is quite a substantial improvement over the current standard. We are now in the process of talking to different institutions in the United States to collect information on additional cohorts to expand our validation efforts.”
In a wide-ranging interview with The ASCO Post, Dr. Li discussed the advantages and potential harms of AI systems in cancer care and how the technology will be used in the future to tailor treatment for individual patients and predict outcomes.
Studying the Tumor Microenvironment
Your AI-based digital pathology study to predict immune checkpoint inhibitor treatment outcomes focused on patients with advanced gastroesophageal cancer. Do you have plans to expand this research analysis into other solid tumors, as well as in blood cancers?
Although our initial focus was on gastroesophageal cancer, we have successfully applied this AI technology to non–small cell lung cancer, and we are exploring its application in a number of other tumor types, including kidney cancer, bladder cancer, and melanoma.4 We haven’t expanded this research yet into blood cancers, because we are concentrating on solid tumors in which we can look not just at the cell morphology, but also at the spatial interaction between different cells within the tumor microenvironment.
Providing Tailored Care for Patients
Based on your research using deep-learning models to predict treatment response and patient outcomes, do you think it is possible to achieve truly personalized care for patients with cancer and experience more cures or durable remissions?
Not at this time. But as these AI-based models with components of precision oncology progress, it will be possible to provide patients with tailored personalized care in the future. In terms of a timeline, we are still in the early days of development of these models. Although many studies are being published, they are still in the discovery phase and mostly with retrospective validation.
The Holy Grail of AI is to achieve personalized treatments for patients with cancer and improve outcomes, which will have the biggest clinical impact.— Ruijiang Li, PhD
Tweet this quote
In our AI-based digital pathology research, we still need to improve the technology. We can go beyond perhaps what we have already identified to get greater high-resolution views into the tumor microenvironment. Once we have achieved that goal, a new predictive model needs to be developed and optimized, which will then be locked down and undergo rigorous prospective validation. Finally, the model needs to be deployed either as a laboratory developed test or an FDA-approved test before the technology can reach patients.
Optimistically, I would put the timeline at about 3 to 5 years. That’s the hope.
Transforming Oncology Care
Please talk about the potential for AI to transform oncology care; what are some of the downsides or potential harmful impacts of this technology on cancer care, especially as it relates to disparities in care?
The Holy Grail of AI is to achieve personalized treatments for patients with cancer and improve outcomes, which will have the biggest clinical impact. But these advancements present a double-edged sword. This technology does pose unique ethical and legal considerations that may limit their broad application and reproducibility, because many AI models are trained with datasets that disproportionately exclude underrepresented groups, so there is inherent bias in these programs, and that’s a problem.
For example, if we develop a mammogram cancer detection AI and the model is trained on data from mostly Caucasian women, the model may not generalize well to patients from other racial and ethnic groups. Biased sampling may lead to poor model performance, inaccurate predictions, and even potential harm. Studies have demonstrated the risk for racial bias in AI algorithms, and we must be cognizant of such bias and figure out ways to reduce it.
However, when used correctly, this technology can decrease health-care disparities by improving diagnostic accuracy and supporting precision medicine. AI can increase access to high-quality care and be deployed to diverse populations in both rural and urban areas of the country, and provide the same high-quality care found in large academic medical centers. That is the benefit of these models.
AI can also automate some of the mundane tasks now performed by physicians, including helping with administrative tasks, such as electronic medical record charting and answering inbox messages, giving them more time to spend with patients, and helping reduce physician burnout.
Resolving the Medical and Ethical Dilemmas of AI
A study by researchers at Dana-Farber Cancer Institute offers an early glimpse of oncologists’ views on the ethical and legal implications of AI tools in oncology care. According to the study results, nearly 85% of respondents said that oncologists should be able to explain how AI models work, but only 23% thought patients need the same level of understanding when considering a treatment; and over 81% agreed that patients should consent to AI use for cancer treatment decisions. When the survey asked oncologists what they would do if an AI system selected a treatment regimen different from their recommendation, 37% of respondents said they would present both options and let the patient decide. When asked who has responsibility for medical or legal problems arising from AI use, 91% of the respondents said AI developers.5
Clearly, oncologists have concerns about who is legally responsible if an AI-recommended treatment results in harm to a patient. In your research, how will these medical and legal issues be resolved before the technology can be used in the clinical care of patients?
This study raises two issues. One is the ability to explain to patients how AI models work, especially when they are used in high-stakes applications, such as treatment decision-making. I absolutely agree with the survey respondents on the potential ethical implications of AI in cancer care, including explainability, or interpretability, and patient consent.
And two, determining who is legally responsible for an AI-recommended treatment will be complicated, partly because it can be challenging to ascertain the cause of patient harm. Depending on the specific circumstance, the responsibility will likely be shared among multiple stakeholders, including health-care practitioners, AI developers and vendors, as well as hospitals and clinics that implement AI systems.
I don’t believe the technology will add another layer of burden for patients, as the decisions on patient care should be made by health-care professionals [in partnership with their patients].
There is no legal regulatory framework yet in place for dealing with these issues, because currently AI systems are being used as an adjunct or support tool and not for primary diagnosis and treatment decision-making. In the end, it is the oncologist who signs off on a patient’s treatment plan.
Adapting the Technology to Clinical and Research Needs
Are you optimistic about the potential for this technology to improve oncology care?
Yes, I am optimistic about the future use of AI in cancer care. The technology will evolve and adapt to clinical and research needs. Although physicians may struggle with how best to incorporate these AI systems into clinical care, we will be able to take a trove of data from routine cancer care and successfully “teach” the AI algorithm to create a personalized treatment approach for patients that considers prognosis and treatment response prediction.
This technology needs to undergo rigorous testing and real-world evaluation before it can reliably weigh-in and make decisions on a patient’s health. Progress is happening fast.
DISCLOSURE: Dr. Li has no financial conflicts of interest to declare.
REFERENCES
1. Schwartz WB, Patil RS, Szolovits P: Artificial intelligence in medicine. N Engl J Med 316:685-688, 1987.
2. Sinha S, Vegesna R, Mukherjee S, et al: PERCEPTION predicts patient response and resistance to treatment using single-cell transcriptomics of their tumors. Nat Cancer 5:938-952, 2024.
3. Eweje F, Li Z, Gopaulchan M, et al: Use of artificial intelligence-based digital pathology to predict outcomes for immune checkpoint inhibitor therapy in advanced gastroesophageal cancer. 2024 ASCO Annual Meeting. Abstract 4013. Presented May 29, 2024.
4. Li Z, Li Y, Xiang J, et al: AI-enabled virtual spatial proteomics from histopathology for interpretable biomarker discovery in lung cancer. Nat Med 32:231–244, 2026.
5. Hantel A, Walsh TP, Marron JM, et al: Perspectives on oncologists on the ethical implications of using artificial intelligence for oncology care. JAMA Netw Open 7:e244077, 2024.
