The Evolution of Artificial Intelligence in Oncology: Impact on Trials, Workflows, and Outcomes

Artificial intelligence (AI) has been utilized in a variety of capacities to fundamentally transform oncology care since its inception. Technology that initially started out as a “hallucination-prone punch line” per Matthew Matasar, MD, chief of the Division of Blood Disorders at Rutgers Cancer Institute and Cancer Survivorship editorial advisory board member for ONCOLOGY, has since evolved into an “increasingly serious” facet of oncology care with a myriad of applications.

The National Cancer Institute (NCI) regards AI as an “unprecedented opportunity” to further understand cancer and improve care for patients with this disease.¹ Focusing on 3 areas that have developed in recent years, including training for AI models, hardware upgrades, and access to large data sets in imaging, genomics, and others, the NCI suggests that the convergence of these aspects has led to “promising new applications” of these tools in cancer research.

The NCI also highlights a greater need to validate AI and machine learning technologies in clinical practice and to advance “explainable” AI to better integrate these technologies into workflows. Additionally, the NCI spotlights a potential for these models to inaccurately represent the broader medical population and perpetuate medical bias with data that are incomplete or not appropriately diverse. In doing so, they acknowledge a need to adopt standards for the development of these technologies to limit bias and retain reproducibility.

For the journal’s 40th anniversary, ONCOLOGY spoke with 3 AI-savvy oncologists: Matasar; Arturo Loaiza-Bonilla, MD, MSEd, FACP, systemwide chief of hematology and oncology at St Luke’s University Health Network; and Patrick Borgen, MD, chair in the Department of Surgery at Maimonides Medical Center, about their insights into the field of AI and its application to oncology. The trio discussed topics such as specialized large language models (LLMs), AI-assisted CT scans for early detection, and foundational models democratizing pathology, among others.

Exploring the Application of AI for Specialized Use in Clinical Workflows

The field of oncology broadened and developed its use of AI, particularly for documentation processes, which frees up clinicians to more directly engage with patients. Matasar stated, “We are at an inflection point with the arrival of AI as a tool in advancing our work in oncology…[with] specialized LLMs…, proliferating companies aimed at streamlining clinical research processes [and] AI-enabled clinical charting capabilities. The last year has seen a tremendous acceleration of early adoption, and yet we have only scratched the surface.”

Regarding LLMs, Matasar highlighted OpenEvidence, an LLM that purports to draw from journals such as the New England Journal of Medicine and the JAMA Network, as well as NCCN guidelines, to deliver AI-generated insights for clinicians.² Their mission is doctor-centered and markets itself as an “AI copilot” to help inform high-stakes, point-of-care decisions. Moreover, this application claims to have supported more than 100 million clinical consultations from US clinicians and is also Health Insurance Portability and Accountability Act compliant and System and Organization Controls
2 Type II compliant per the American Institute of Certified Public Accountants.

Additionally, in response to how AI is incorporating workflows, Loaiza-Bonilla expressed that it could be used to optimize them and function as a sort of “connective tissue” to accelerate clinical processes.

“From the clinical aspect, AI is becoming more like a connective tissue in the oncology continuum,” he explained. “It’s not about, ‘Should I adopt it?’ It’s more, ‘How do I adapt my workflow to the advent of AI to the benefit of my practice and my patients?’ We have seen agentic AI supporting tumor boards [and] clinical trial matching, and we do pathway adherence, care coordination, and embedding with guidelines [from the American Society of Clinical Oncology] and NCCN.”

Agentic AI, which utilizes AI orchestration to coordinate machine learning models to perform tasks, uses LLMs to build upon existing generative AI techniques.³ Furthermore, setting it apart from other generative AI models is its capacity to use generated content, calling upon external tools to complete complex tasks autonomously. Loaiza-Bonilla further stated that this technology is not intended to replace clinicians but to remove friction points that arise during routine clinical practice.

“The most meaningful deployments are not going to be flashy or [announced in] press releases. It is going to be if we can reduce documentation burden, if we can accelerate triage, if we can get access to trials faster and help teams execute—that is going to be the real value that we’re going to see in deployments of these tools,” Loaiza-Bonilla explained.

Supplementing Radiology Scans for Earlier Cancer Detection

In alignment with 2 major trends, mammography and early detection in lung cancer screening, AI is being used to better identify early-stage cancers in patients. Loaiza-Bonilla explained that the use of AI in mammography may help to cut down on the time and effort associated with the procedure.

“We are now using AI to assist with [mammography], to decrease the amount of double review on those mammograms, and it’s also being run on clinical trials,” he said. “We have data from Europe, but now there are several efforts in the US as well, because our population is different in terms of breast density and beyond. I’m excited to see the results of those algorithms as they get deployed.”

Beyond screening for lung and breast cancer, Loaiza-Bonilla touched upon an initiative led by Friends of Cancer Research to establish criteria for integrating AI models into RECIST assessments to better measure lesions in a more scalable way.⁴ Moreover, he touched upon findings from the PANORAMA trial, published in The Lancet Oncology, which highlighted the capacity for AI-assisted CT scans to detect early-stage pancreatic cancer, given that it is typically identified in later-stage indications.⁵

He concluded by emphasizing a desire to see these models deployed in real-world practice, stating that many radiologic models have been localized to a small set of use cases.

“What I hope for is to start using those models in real practice. As of now, many of the models that we use in radiology have been localized to those couple of use cases and for the emergency department, such as finding fractures, but now, with oncology being one of those very high-stakes diseases, we have an opportunity here,” he said.

Facilitating Generalizability in Cardio-Oncology Through Foundational Models

According to Loaiza-Bonilla, foundational models will help facilitate clinical processes in cardio-oncology, particularly for electrocardiograms (ECGs). Foundational models, which are data sets encompassing up to millions of labeled tests, can be built for generalizability to specific use cases.

“In this case, it could be the detection of arrhythmias or risk factors such as QT prolongation in patients taking [tyrosine kinase inhibitors] or [adverse] effects that are emerging, particularly for patients who may have predispositions such as cardiovascular disease,” he explained.

Outlining the state of the field today, he identified a trend among companies seeking to validate these foundational models with the FDA to democratize their use in cardio-oncology. He anticipates that the ubiquitous nature of ECGs and their ease of use will permit expedited FDA approval for use as risk-stratification tools. These models, along with other data-driven algorithms, may help “move the needle” in cardio-oncology when integrated into electronic medical records.

Advancements in AI-Assisted Breast Cancer Screening

At the 2025 San Antonio Breast Cancer Symposium (SABCS), a transformer-
based architecture demonstrated the ability to predict the risk of recurrence among patients previously treated for breast cancer.Moreover, Loaiza-Bonilla expressed that this tool could be used to optimize recurrent response scores and identify patients with rarer disease subtypes.

“[The architecture] can optimize those recurrent response scores that we are seeing from different vendors in a more meaningful way and find those niche patients [who] need either more screening or early detection of other things, such as germline alterations or risk factors, when we use any combination with other data sets,” he said.

Moreover, he highlighted that leveraging the use of AI in radiomics may help to ascertain whether a patient would be likely to experience benefit from treatments, such as endocrine therapy, simply by observing a CT scan. Highlighting findings from the phase 3 TAILORx trial (NCT00310180), he explained that the use of AI tools has become a “major validation point” for models being deployed or undergoing testing for use in the clinic.⁶

Findings presented at SABCS revealed that the multimodal model, which integrates imagining, clinical, and expanded molecular models, displayed a strong prognostic performance for overall and late disease recurrence, which was superior to the Oncotype DX 21-gene recurrence score alone. It exhibited statistically significant and clinically relevant prognostic stratification in low– and high–genomic-risk groups.

Moreover, Borgen stated that AI could be used to better screen patients and read breast imaging studies. Despite being novel technology requiring further validation, he envisions its use as a cost-effective and more efficient tool to enhance oncology practice.

“We’ve already seen data that AI does as good or better a job at reading breast imaging studies. Certainly in the screening setting, AI is now able to read 3D or tomosynthesis mammograms, which it wasn’t able to do in the past,” he explained. “We’re waiting for them to add a feature that will compare to old x-rays, but at some point, it’s going to replace some humans…. It’s not ready for prime time, not ready to invest in a company, but the preliminary data are compelling.”

Democratizing Pathologic Oncology Through Foundational Models

Contextualized by a shortage of pathologists in low- and middle-income countries, as well as certain sections of the rural US he dubbed “deserts of cancer care,” Loaiza-Bonilla believes that the answer may lie in foundational models to expedite immediate readouts to identify patients in need of further testing or treatment in the absence of readily available pathologists.

“We are, at this moment, very short globally, not only in low- and middle-income countries, but in counties in the US, [on] pathologists [who are] readily available to review slides and give us support. One of the things that we’ve seen now is with these big foundational models of millions of slides being taken, whole slide imaging pictures, we can leverage those models to look at just the [hematoxylin-eosin] stain,” he explained.

He further expressed that in using these democratizing tools at scale, the data can be uploaded via cloud and sent to a local computer, enabling the assessment of initial screening to be done remotely from anywhere. Moreover, with these models, one could identify risk for potential biomarkers even before a pathologist performs their own read of the slide.

“[You] simply take the picture and deploy and say, ‘Hey, this may have an EGFR mutation or this sample may be [estrogen receptor]/[progesterone receptor] positive or have PD-L1 positivity.’ It can be consequential in the use of the pathology,” he stated. “[Multidisciplinary] efforts need to happen across [practices] to implement this, but the promise is tremendous.”

Borgen added that he acknowledged a scenario where AI image analysis replaces studies in the genomic profiling space. “In pathology…these platforms learn incredibly quickly and can be instructed,” he explained. “One question is: ‘Of all the tests that are out there in the genomic profiling space, is image analysis using AI going to replace those studies?’ It’s certainly a possibility.”

2026: The Year of AI Industrialization in Oncology

According to Loaiza-Bonilla, 2026 will mark the year of industrialization in AI. Moving beyond infrastructure-embedded services, health and pharmaceutical companies are investing in the licensing and training of foundational models on their own proprietary data sets, with the intent of accelerating the process of drug discovery to patient deployment.

“We are seeing discovery cycles being compressed from years to sometimes months and, in a few cases, weeks. That is how we are going to see an improvement,” he explained.

Agentic AI, which he stated would be used for the “final determination of steps” in this process, purports to help filter through publications and compile “raw news into informed decisions.” Specifically, it serves as a means of compiling information for niche and technical spaces and filtering it to deliver keen insights that might not be readily available from a sweep of mainstream news outlets.

Moreover, despite the emergence of AI-centric platforms, Loaiza-Bonilla stressed that oversight must be implemented in the development of these systems to align them with the evidence-based literature and combat misinformation.

“The opportunities are huge, but they need to be aligned with evidence, with safety, with workflows, and, of course, we want to avoid any misinformation,” he noted. “[There] has to be clinical involvement, governance, and some guardrails, at least with specialty-driven pathways. Hopefully, that translates into what I feel are the best use cases, which are speed, access, and execution.”

Among platforms for patients that he mentioned were ChatGPT Health and Claude. ChatGPT Health claims to be a dedicated health experience, with the functionality of ChatGPT plus the ability to connect information such as medical records and wellness apps to help patients with their health-related decisions.⁷ Claude draws from a “family” of LLMs and offers 3 models that vary by size and complexity, with the smallest, Haiku 4.5, offering quicker insights, and the largest, Opus 4.6, offering complex analysis, deep research, and document generation.^8,9

Loaiza-Bonilla concluded by explaining that for AI models to become a foundational component of his practice, they must adhere to his best use cases.

“Did [AI] bring therapies to patients faster, or were they very informed? Can we get access to clinical trials and guideline-concordant care faster? Did my team and [I] spend more time practicing medicine and less time fighting systems? If AI helped us do those things, [it has earned its] place in oncology,” he explained. “That’s how it is going to [move] from a peripheral technology to an interesting foundational component of our workflows.”

References

1. Artificial intelligence (AI) and cancer. National Cancer Institute. Accessed February 10, 2026. https://tinyurl.com/48jdj42y
2. About. OpenEvidence. Accessed February 9, 2026. https://tinyurl.com/3yzx2bcx
3. Agentic.ai home page. Accessed February 9, 2026. https://tinyurl.com/yc5jrm5f
4. Ai.RECIST Project: Artificial Intelligence-Enabled Response Evaluation Criteria in Solid Tumors Project. Friends of Cancer Research. Accessed February 9, 2026. https://tinyurl.com/rx26459z
5. Alves N, Schuurmans M, Rutkowski D, et al. Artificial intelligence and radiologists in pancreatic cancer detection using standard of care CT scans (PANORAMA): an international, paired, non-inferiority, confirmatory, observational study. Lancet Oncol. 2026;27(1):116-124. doi:10.1016/S1470-2045(25)00567-4
6. Sparano JA, Lama N, Gray RJ, et al. Multimodal artificial intelligence (AI) models integrating image, clinical, and molecular data for predicting early and late breast cancer recurrence in TAILORx. Presented at: 2025 San Antonio Breast Cancer Symposium; December 9-12, 2025; San Antonio, TX. Abstract 1601.
7. Introducing ChatGPT Health. OpenAI. January 7, 2026. Accessed February 9, 2026. https://tinyurl.com/3rux7brr
8. What is Claude AI? IBM. Updated February 3, 2026. Accessed February 9, 2026. https://tinyurl.com/2rwv9ar4
9. Overview. Claude. Accessed February 9, 2026. https://tinyurl.com/5n82a9tx