Beyond geometry: How biomarkers and AI are redefining personalization in radiation oncology

For decades, radiation therapy was a question of millimeters, with health care providers using ever more sophisticated imaging and planning to hit tumors accurately, while sparing nearby organs at risk. Now, the discipline is on the cusp of a shift in which biological insight is beginning to complement, and in some cases challenge, the discipline’s traditional focus on geometric precision.

The transition was the theme of a recent GE HealthCare-sponsored symposium at ESTRO in Stockholm, Biomarker-guided therapy: Advancing precision oncology^[i], where three academic radiation oncology researchers explained how imaging biomarkers, hidden biological variables and Artificial Intelligence (AI) could help to make radiotherapy more adaptive and more individualized.

In a lunchtime session held on May 18 at the European Society for Radiotherapy and Oncology (ESTRO) Congress and moderated by Ilya Gipp, MD, PhD, GE HealthCare’s Chief Medical Officer for Oncology, three experts outlined their vision of the future for radiation oncology. One theme was consistent: Progress will depend on understanding tumor biology, response variability and clinically useful biomarkers, rather than simply refining dose delivery alone.

Laure Marignol, Professor in Radiation Biology at Trinity College Dublin, said that the discipline has “largely mastered” geometry, imaging guidance and anatomical targeting, and that future gains in outcomes would come from identifying the biological factors that determine why patients respond differently to the same treatment. “We really know very well where to treat,” she said. “The next challenge is understanding who we are truly treating biologically,” she added.

Imaging biomarkers: cautious optimism

Claudio Fiorino, a senior medical physicist at San Raffaele Scientific Institute in Milan, highlighted the potential of Quantitative Imaging Biomarkers (QIBs), the numerical features extracted from routine imaging that correlate with biology or outcomes. Since radiotherapy already depends on imaging such as CT, MRI and PET scans throughout planning and treatment, these data are readily available sources of biomarkers that could potentially predict toxicity, response and outcomes without the need for new tests, explained the academic. “Medical images are available without additional cost for all patients,” said Fiorino, and “they can capture a lot of characteristics.”

Fiorino gave examples, explaining how San Raffaele clinicians analyzed CT densitometry measurements derived from “simple and available” planning CT scans of more than 1,500 patients for evidence that cardiac calcification scores predict cardiac events. They also analyzed lung density measures and fat percentages from CT scans to predict patient outcomes.

The academic sounded a note of caution about the clinical readiness of QIBs, warning that the “so-called variability” of the data, or radiomic features, would be a challenge.  Fiorino explained how uncertainty is introduced at every stage of the imaging workflow, from acquisition to feature computation. He cited an influential radiomics survival model^[ii], that later inspired another study^[iii] driven largely by tumor volume, to establish a set of safeguards designed to improve and support current radiomic methodologies through detailed analysis of a radiomic signature.  Fiorino ended on an optimistic note, reminding delegates that “we have examples of robust biomarkers [that] we can trust.”

Hidden biology: It matters

While Fiorino drew attention to imaging’s untapped potential, Marignol focused on what standard workflows still fail to capture. The Dublin-based academic said that some of the most important drivers of response remain biologically hidden, including hypoxia, DNA damage response, immune activity, inflammation, host factors and chromosomal instability. “Patients may receive the same dose, but they are not biologically equivalent,” she said.

Marignol warned that ignoring biology results in standardized treatment decisions, which masks important differences between patients and populations. Citing the example of sex as an underused variable, the academic noted that around 30% of male tumors have lost the Y chromosome, a phenomenon that might affect repair, survival and immune response. Yet in a review of 321 radiation oncology papers, fewer than half included sex in the analysis, while only two separated outcomes by sex, said Marignol. “These hidden variables matter,” she added. “Precision radiotherapy without biology risks inequity.”

Heterogeneity, AI and the next frontier

Robert Jeraj, Professor of Medical Physics, Human Oncology, Radiology and Biomedical Engineering at the University of Wisconsin-Madison, leaned further into the problem of tumor heterogeneity. He shared examples of spatial heterogeneity within single tumors (intra-lesion heterogeneity) and variable responses among multiple lesions (inter-lesion heterogeneity) in metastatic disease. These phenomena are important, since they undermine the one-size-fits-all assumptions about tumor response, and raise the possibility of adapting treatment not just between patients, but within the tumor (e.g., biologically-targeted radiotherapy, or “dose painting”) or between lesions (e.g., targeting oligo-resistance) and over time. 

But QIBs have a role to play in solving the problem, said the academic, since the modality is inherently spatio-temporal, allowing clinicians to understand how a tumor behaves in space, and how its behavior changes during treatment. “Quantitative imaging allows us to uncover this heterogeneity,” said Jeraj. However, the academic warned that identifying heterogeneity was only one part of the problem, since a single form of treatment would be insufficient. “Combining systemic and local therapy, such as radiation therapy, is key,” he said.

AI becomes essential in this highly complex setting, added Jeraj. Advanced analytics could identify resistant lesions, support lesion-level decision-making and extend the role of radiotherapy in advanced disease.  “Humans cannot manage this complexity alone,” he said. In a light-hearted moment, Gipp urged the responsible use of AI in biomarker-guided radiotherapy, reminding delegates of the pitfalls of modern dating apps. “The algorithms promise highly personalized matches,” said the moderator. “Yet somehow many end up with an excessive dose of unwanted attention, and relationships developing serious toxicities,” he added, to laughter from the room.