During her Presidential address at the 2025 ASCO Annual Meeting, Robin T. Zon, MD, FACP, FASCO, assessed how artificial intelligence (AI) is driving knowledge into action in the field of oncology, and acknowledged that “we are now at the crossroads of long-imagined possibilities and actionable solutions” because of this technology.
Robin T. Zon, MD, FACP, FASCO
“AI is improving our ability to acquire, assimilate, and organize knowledge,” said Dr. Zon. “Did you know there is at least one new medical article published every 26 seconds? You would have to read 5,000 articles a day to keep up. It’s not surprising that simply trying to stay current is one of the causes of burnout, but AI is helping dramatically.”1 Dr. Zon also agreed with Harvard Business School Professor Karim R. Lakhani, PhD, that “AI won’t replace humans—but humans with AI will replace humans without AI.”2
In this Special Report, we examine how AI is rapidly transforming oncology care by enhancing diagnostic precision in pathology, accelerating clinical trial matching, and enabling personalized treatment strategies—all with the goal of improving patient outcomes. We also explore the ways in which AI is reshaping how patients with cancer receive health information, shifting more control of care to patients, and underscoring the need for critical health AI literacy to evaluate outputs for bias and accuracy.
The report is divided into five sections: Reimagining Cancer Pathology; AI-Powered Clinical Trial Matching; AI-Driven Clinical Decision Support; Critical AI Health Literacy to Empower Patients; and Personalizing Clinical Care: A Look Into the Future.
Gabriele Campanella, PhD
Ethan Cerami, PhD
Our experts include Gabriele Campanella, PhD, Assistant Professor of the Windreich Department of Artificial Intelligence and Human Health at the Icahn School of Medicine at Mount Sinai in New York City; Ethan Cerami, PhD, Director of the Knowledge Systems Group and Principal Scientist, Department of Data Science at Dana-Farber Cancer Institute; Ravi B. Parikh, MD, MPP, FACP, Associate Professor, Department of Hematology and Medical Oncology at Emory University School of Medicine, Medical Director of the Winship Data and Technology Applications Shared Resource at Winship Cancer Institute, and Editor-in-Chief of ASCO AI in Oncology; Hugo Campos, Strategic Advisory Board Member, Computational Precision Health Strategic Advisory Board at the University of California, Berkeley, and the University of California, San Francisco; and Ruijiang Li, PhD, Associate Professor of Radiation Oncology, Division of the AI for Precision Oncology Laboratory at Stanford University School of Medicine.
Ravi B. Parikh, MD, MPP, FACP
Hugo Campos
Ruijiang Li, PhD
Reimagining Cancer Pathology: Gabriele Campanella, PhD
Over the past 65 years, AI in cancer pathology has evolved from early digital imaging efforts in the 1960s to 1980s, to whole-slide scanners enabling the digitization of glass slides in the 1990s to 2000s, and to the widespread adoption of convolutional neural networks in the 2010s. Today, deep-learning–driven diagnostic tools are increasingly used to assist human expertise in the clinical setting.
A recent study by Dr. Campanella and colleagues shows how AI can enhance tumor sample analysis and improve clinical care. The study investigated EGFR AI Genomic Lung Evaluation (EAGLE), a tool designed to predict EGFR mutational status from the diagnostic biopsies of patients with lung adenocarcinoma using standard hematoxylin and eosin-stained pathology slides. Results showed that AI could reliably detect EGFR mutations, potentially reducing the need for rapid genetic testing by more than 40%,3 which could streamline clinical decision-making, preserve limited tissue for more comprehensive sequencing if needed, and accelerate patients’ access to targeted therapies, according to Dr. Campanella.
In this interview with The ASCO Post, Dr. Campanella discusses his study findings, whether AI may merge radiology and pathology into a single specialty, and its potential to help alleviate the shrinking pathology workforce.
What are the next steps in your research on integrating an AI foundation model into digital pathology, and what is the process for U.S. Food and Drug Administration (FDA) approval in this rapidly evolving area?
I would say that, compared with radiology, there are far fewer FDA approvals of AI systems in pathology, so the regulatory landscape remains uncertain. I do agree that the scope of the models will be important in determining the regulatory path. For systems that merely replicate a known biomarker, as is the case with EAGLE, regulation may fall under the laboratory-developed test framework. By contrast, for completely novel biomarkers, these models may require full FDA approval.
In terms of our research, we are continuing to collect data in our EAGLE study and plan to broaden the system’s capabilities to detect additional cancer biomarkers to support potential clinical deployment. We are also evaluating the technology’s impact in low-resource settings, where access to next-generation sequencing is limited and this AI foundation model could make a meaningful difference in clinical care.
A paper by Sala et al contends that AI is reshaping the foundations of diagnostic medicine and blurring the boundaries between image interpretation and tissue analysis, providing a rationale for merging imaging and pathology into a unified “diagnostic medicine” specialty that could enable more accurate, timely, and personalized care.4 Do you see this proposal as the future of diagnostic medicine? Could these fields merge?
I did not read this commentary, but in my view, I do not see a merging of these fields anytime soon. There is a reason these two fields are separate. There is hope that AI will supercharge the capabilities of radiology, so that in certain circumstances, invasive procedures like tissue biopsies may not be necessary. However, I don’t see these two pillars of medicine merging, at least not for decades.
Years of experience in pathology training will remain essential, even in this fast-evolving era of AI-enhanced tools.
How do you envision the responsible use of AI-powered pathology tools in clinical oncology in the future?
The responsible approach is to leverage technologies that have been proven safe and effective to improve health care and patient outcomes. Some AI-based technologies have reached clinical-grade performance, which is not a blank check for their liberal use. However, if a system has been shown to improve care, it should be used. At the same time, there should be a high burden of proof on developers to ensure safety and integration into real-world clinical workflows.
We have a workforce that is aging and retiring just as cancer incidence is increasing, and not enough young pathologists are entering the field to replace them. In the years ahead, cancer care will become even more complex, and the need for precision oncology will continue to grow. Artificial
intelligence can serve as an additional tool to help address this major challenge. It is not the only solution, but it is one way to increase efficiency in the clinical workflow.
However, it is important to keep in mind that this human-centered technology is an assistive tool, not a replacement for human clinicians.
AI-Powered Clinical Trial Matching: Ethan Cerami, PhD
Cancer clinical trial enrollment remains abysmally low, at between 5% and 7% of adult patients.5 A recent review by Dr. Cerami and colleagues of developments in AI applications for oncology clinical trials found that the technology may play a role across the clinical trial research spectrum, including AI-powered drug design, optimization of trial execution, trial matching, and eligibility prescreening.6
Here, Dr. Cerami discusses how AI is advancing clinical trial matching and design, as well as the cautions ahead.
Please talk about the findings from your research. How might AI accelerate cancer clinical trial matching?
Much of our work focuses on developing computational systems designed to match patients to clinical trials. These systems existed before AI but are now improving with advances in AI, including large language models (LLMs).
From an oncologist’s perspective, we have extensive patient data, and there are hundreds of clinical trials that may be open across sites. A key challenge is rapidly identifying which patients match specific trials. There is a clear need for computational solutions that can perform patient-trial matching more efficiently.
At Dana-Farber, we have developed two open-source computational platforms to match patients to clinical trials. The first, MatchMiner Genomics (https://matchminer.org), is designed to match patient-specific genomic profiles to precision cancer medicine trials. The second, MatchMiner-AI, is our new AI platform that leverages LLMs to match patients to trials based on their full medical record.
Both systems are fully integrated into our electronic medical record system, allowing oncologists to quickly access patient information and identify the most relevant clinical trial options.
Artificial Intelligence in Oncology Care: AI is rapidly transforming oncology care by enhancing diagnostic precision in pathology, accelerating clinical trial matching, and enabling personalized treatment strategies—all with the goal of improving patient outcomes. This technology has placed us “at the crossroads of long-imagined possibilities and actionable solutions.”
How can deep learning AI models such as AlphaMissense and AlphaGenome, which interpret genetic variations, aid in drug design and identification? If successful, could AI help more precisely match patients to clinical trials with effective therapies that improve outcomes?
Cancer is caused by molecular changes at the DNA level, and new AI tools, including AlphaMissense and AlphaGenome, are designed to interpret these alterations and predict which are likely to be functional or linked to cancer. This represents a powerful development that can help identify new molecular targets, ultimately driving drug discovery.
AI models can, therefore, potentially accelerate the identification of novel treatment candidates, with the goal of developing new therapeutics that improve patients’ lives.
In your study, you warn that caution is necessary when using AI to predict clinical trial success. What are your concerns about this technology?
There is considerable hype surrounding AI. However, the medical community already has a well-honed tool to cut through that hype: clinical trials. Any AI-designed drug or intervention will undergo the same rigorous evaluation of patient impact as non-AI interventions.
One shining example is the use of AI in mammography. Multiple companies have developed AI-assisted tools that use advanced algorithms to help radiologists to detect breast cancer, enhancing accuracy and improving workflow efficiency.
Large-scale clinical trials underway in the United States and Europe are objectively evaluating these tools, and results to date suggest a measurable positive impact on patients.
What are the future directions of this technology in cancer care, and what is the expected timeframe for AI-driven tools to be fully integrated into clinical care for patients with cancer?
There are certain areas of clinical care, such as mammography and pathology, where AI-driven tools have already made significant improvements in screening and diagnosis.
AI-driven clinical trial matching is another area likely to become more widespread across cancer centers in the near future, and improving clinical trial enrollment is an obvious need. My hope is that, once these AI systems are in place, we can maximize patient participation in clinical trials. This would enhance our ability to run and complete trials more efficiently and accelerate the evaluation of whether specific therapeutic interventions benefit particular patients.
Looking ahead, the use of generative AI in other clinical areas—for example, to provide treatment recommendations—remains largely a research question. Significant challenges and regulatory hurdles must be addressed before such systems are widely used in clinical care.
AI-Driven Clinical Decision Support: Ravi B. Parikh, MD, MPP, FACP
Dr. Parikh leads the Human-Algorithm Collaboration Laboratory at Emory University, where he is focused on developing trustworthy AI models that enhance clinical decision-making in oncology. In our conversation, Dr. Parikh talked about how AI is assisting in treatment decision-making, the ethical and legal considerations of this technology, and who should be held responsible if an AI-assisted treatment decision goes wrong.
How do you envision AI supporting clinical decision-making, and what progress have you made in your research?
When I think about how AI will enhance clinical decision-making, it is largely by making treatment selection more individualized for patients. We typically rely on clinical trial data to inform treatment decisions. However, it is estimated that between 30% and 40% of patients would not qualify for a clinical trial—so how do we generalize the results from clinical trials for that segment of the patient population?
Our laboratory work has focused on the concept of a “digital twin,” a computational representation of a real-world patient. Using AI-based methods, we can not only match patients to therapies with the same accuracy as in a clinical trial, but we can also simulate the early trajectory of a patient’s illness.
For example, we may be able to simulate the impact of different treatment sequences, which are often tested in clinical trials. We may also be able to simulate drug effectiveness by emulating trials in populations that have been underrepresented in trials.
Through this work, we envision moving beyond isolated tissue-based or circulating tumor DNA (ctDNA) biomarkers to more comprehensive predictors of treatment response. This approach could provide a more realistic estimate of how likely a patient is to benefit from a given therapy—shifting from average treatment effects observed in trials to individualized patient outcomes, which is the ultimate goal of the digital twin.
How far along is your digital twin research, and when might it be available in the clinical setting?
Our research focuses on developing digital twins for cancer-risk prediction. We are building predictive models in breast and prostate cancers that integrate patient-specific data, including genomics, imaging, and clinical history, to simulate tumor evolution and predict treatment response. The next challenge is to validate these tools so they can be approved by the FDA. To date, predictive digital twins have not yet been approved as a distinct, standardized category of medical device.
Currently, we view our digital twin tools as proof of concept and potentially useful in research settings. However, for clinicians to trust these predictive treatment tools over more established methods, they will need to be evaluated in a retrospective or high-quality prospective study and compared with existing biomarkers, such as blood tests and imaging. Once we have a clear sense of how they benchmark against biomarkers used in routine clinical practice, digital twins may become true adjuncts in clinical care. Otherwise, they are likely to remain largely in the research domain.
What are the ethical considerations and associated risks, such as legal and informed consent issues, in using this technology in patient care?
One of the biggest ethical concerns is that digital twins and related biotechnology and bioengineering biomarkers, whether pathology-based, biology-based, genomic-based, or multimodal, are often built on highly homogeneous datasets, which may be dominated by certain racial or socioeconomic groups. Another concern is that these models may rely on data sources in which patients are unaware that their information is being used.
This remains a controversial issue, and we have not yet determined what level of data sharing absent explicit patient consent is acceptable for the public good.
The second question is: What does it take to make these tools truly trustworthy in the clinical setting? AI-based biomarkers may be more accurate than conventional diagnostic tools, but they inherently suffer from limited explainability, ie, the “black box” problem, which makes it difficult to understand how a system reaches a specific conclusion. We need to make AI models more transparent while maintaining high performance; we are already seeing tension between trustworthiness and actual clinical benefit that must be resolved.
Although the use of AI-powered tools in oncology care remains largely unregulated, this technology is already being used to support treatment decision-making. Who is responsible if an AI-assisted decision leads to an adverse outcome?
The short answer is that liability ultimately rests with the clinician. We are responsible for how these tools are used. As a physician, I may not always know whether an AI-based tool has been integrated into my institution’s health system. We may be willing to accept some degree of liability if these tools provide a broader societal benefit. However, we should be developing liability frameworks that also hold algorithm developers—and potentially the health systems utilizing these tools—accountable if something goes wrong, rather than placing the burden solely on individual physicians.
Critical AI Health Literacy to Empower Patients: Hugo Campos
More than 40 million people worldwide use ChatGPT daily for health-related questions and to navigate complex health-care systems.7 A growing number of patients with cancer, about one-third, are also consulting AI chatbots to interpret lab reports, scans, and symptoms, often before speaking with their oncologists.8
In this interview, Mr. Campos discusses how patient-directed AI can help patients influence the health-care systems that shape their care, shifting control from institutional priorities to the patient, and how critical AI health literacy may enable more equitable, patient-centered treatment decision-making.
You and your colleague, Liz Salmi, AS, have written about the need for patients to adopt critical AI health literacy to achieve empowerment.9 How do you define critical AI health literacy?
We define the concept as the “ability to strategically use AI tools to analyze health-determining factors and power structures; to evaluate AI outputs for bias and institutional alignment; and to take informed action to advance individual and collective health goals through algorithmic resistance to systems that prioritize institutional priorities over patient values.”
At the center of this definition is the idea that patients need to understand the difference between institutional and patient-directed AI. Institutional AI serves organizational priorities, such as compliance, efficiency, and cost containment, whereas patient directed AI serves the patient.
AI-powered tools in clinical oncology are becoming increasingly ubiquitous. In a recent survey of cancer survivors’ perceptions of AI use in oncology, 44% reported feeling scared of the technology.10 How can patients with cancer develop AI health literacy to become more active participants in their care?
That 44% figure is understandable—I would have expected an even higher percentage. However, people who are afraid of AI need to recognize that it is already embedded in oncology care, from imaging and pathology to treatment planning and insurance decisions. It is not a question of whether patients will encounter this technology, but rather whether they will have the skills to engage with it in ways that align with their health-care goals.
The goal of critical AI health literacy is not simply to help patients understand how AI is being used, but to help them develop the awareness and skills to use AI as a thinking partner in ways that expand agency, question assumptions, and act more effectively in their care.
What responsibility does the oncology team have to ensure that patients have the tools they need to use AI to gain greater control over their disease and participate in treatment decision-making?
The oncology team has a fundamental responsibility that begins with transparency. Clinicians need to be open with patients about where AI is being used in their care. If an algorithm helped interpret a patient’s scan, flag a treatment protocol, or prioritize a referral, the patient deserves to know.Beyond transparency, care teams should actively encourage patients to use AI tools as part of their care preparation, rather than discouraging them.
We need to reframe AI from something inherently threatening to something constructive and collaborative. AI is already widely used, and it is disingenuous to discourage patients from using it when doctors, health systems, and insurers rely on it every day.
Personalizing Clinical Care: A Look Into the Future: Ruijiang Li, PhD
An AI model developed by Dr. Li and colleagues at Stanford University School of Medicine has demonstrated accuracy in predicting disease outcomes, including melanoma relapse, pan-cancer prognosis, and immunotherapy response in patients with lung and gastroesophageal cancers.11 Dubbed MUSK (multimodal transformer with unified masked modeling), the tool is a vision-language foundation model designed to leverage large-scale, unlabeled, unpaired image and text data.
In the study, for non–small cell lung cancer, MUSK correctly identified patients who benefited from immunotherapy treatment approximately 77% of the time, compared with about 61% using the standard method based on PD-L1 expression. It also identified individuals with melanoma most likely to relapse within 5 years of their initial treatment with approximately 83% accuracy—about 12% more accurate than predictions generated by other foundation models.11
In this interview, Dr. Li discusses the importance of developing AI-based predictive models, how clinicians can use them to guide treatment, and how this technology may lead to more cancer cures.
What is the significance of your research findings, and how might they change the paradigm of the information clinicians use to guide treatment?
The motivation behind this research was to integrate patient information that is often siloed across specialties. For example, typically, a radiologist will review a scan, a pathologist will study slides under a microscope, and oncologists will base their prognosis on patient-specific factors documented in clinical notes.
This information remains siloed and is not fully integrated into a single foundation model. From a technical standpoint, MUSK represents a novel approach that can leverage large-scale labeled and unpaired image data—for example, a microscope slide of a piece of malignant lung tissue—as well as paired data from clinical notes. Traditional methods often require paired data, which are more limited in practice. By leveraging large-scale unpaired data to train MUSK, we were able to achieve improved performance.
The most valuable clinical applications for this technology will be in treatment decision-making. Currently, clinicians rely on factors such as disease staging and specific genes or proteins to guide decisions, but these approaches are not always accurate. Because treatment selection is a high-stakes decision, MUSK will need to be validated in additional clinical trials before it can be used in practice, and it will also require FDA approval as a high-risk medical device.
Based on your research so far, how will AI transform oncology care? Could it enable more cancer cures or shift more cancers into chronic diseases, and what might this mean for individual patients?
The most impactful application of this technology over the next few years will be the advancement of true precision medicine in oncology. My research focuses on using AI to integrate spatial biology and pathology to uncover novel insights into disease biology and the mechanisms of response and resistance to therapy, moving beyond the traditional one-size-fits-all approach.
That’s in the short term. Over the longer term, the greatest impact will likely be in drug discovery and the development of new medicines that will be more effective against cancer, with the expectation that achieving true precision medicine will improve patient outcomes. However, some cancers will remain particularly challenging to treat, including glioblastoma and pancreatic cancer.
But given a longer timeline, this technology hasthe potential to cure more cancers.
DISCLOSURE: Dr. Zon owns stock in Moderna, Oncolytics Biotech, TG Therapeutics, Select Sector SPDR Health Care, AstraZeneca, CRISPR Therapeutics, McKesson, and Berkshire Hathaway and is a member of Cincinnati Cancer Advisors. Dr. Campanella is a consultant for Daiichi Sankyo. Dr. Cerami, Dr. Parikh and Mr. Campos have no financial conflicts of interest to declare. Dr. Li is the co-founder and scientific advisor for Perception Medicine, Inc.
REFERENCES
1. ASCO Connection: 2025 President’s Address: “Driving Knowledge to Action: Building a Better Future.” June 1, 2025. Available at https://connection.asco.org/do/2025-president-s-address-driving-knowledge-action-building-better-future. Accessed May 6, 2026.
2. Lakhani KR: AI Won’t Replace Humans—But Humans With AI Will Replace Humans Without AI. Harvard Kennedy School, August 4, 2023. Available at https://hbr.org/2023/08/ai-wont-replace-humans-but-humans-with-ai-will-replace-humans-without-ai. Accessed May 6, 2026.
3. Campanella G, Kumar N, Nanda S, et al: Real-world deployment of a fine-tuned pathology foundation model for lung cancer biomarker detection. Nat Med 31, 3004-3019, 2025.
4. Sala E, Goyen M: E. pluribus unum: The diagnostician.EJR-AI 4 100045, 2025.
5. Gong G, Liu J, Pandva S, et al: Clinical trial patient matching: A real-time common data model and artificial intelligence-driven system for semiautomated patient prescreening in cancer clinical trials. JCO Clin Cancer Inform 10: e2500262, 2026.
6. Cerami E, Riaz IB, Kehl KL: AI for clinical trials in oncology. ESMO Real Worl Data Digit Oncol 11:100658, March 2026.
7. Olsen E: 40M users turn to ChatGPT daily for health questions: OpenAI. Healthcaredive, January 6, 2026. Available at https://www.healthcaredive.com/news/40-million-use-chatgpt-health-questions-openai/808861/. Accessed May 6, 2026.
8. Rosenbluth T, Astor M: Empathetic, available, cheap: When A.I. offers what doctors don’t. The New York Times, November 17, 2025. Available at www.nytimes.com/2025/11/16/well/ai-chatbot-doctors-health-care-advice.html. Accessed May 6, 2026.
9. Campos H, Salmi L: Critical AI health literacy as liberation technology: A new skill for patient empowerment. National Academy of Medicine, December 8, 2025. Available at https://nam.edu/perspectives/critical-ai-health-literacy-as-liberation-technology-a-new-skill-for-patient-empowerment/. Accessed May 6, 2026.
10. Fortune E, Newell A, Pink S, et al: Patient perceptions of artificial intelligence in healthcare: Findings from the Cancer Experience Registry. NCCN 2025. Abstract BIO25-023. Presented March 28-30, 2025.
11. Xiang J, Wang X, Zhang X, et al: A vision-language foundation model for precision oncology. Nature 638(8051):769-778, 2025.
