The review, published on 3 September 2026, does not claim that radiomics should replace tissue-based biomarkers. Its central argument is more subtle and, arguably, more clinically useful: radiomics should serve as a spatial context that tells clinicians when a biopsy result can be trusted and when it might be dangerously misleading. Head and neck squamous cell carcinoma, or HNSCC, is an ideal testing ground for this idea. These tumors arise in anatomically complex regions — the larynx, oropharynx, oral cavity, hypopharynx — where treatment decisions carry enormous consequences for speech, swallowing, and quality of life, and where the stakes of both overtreatment and undertreatment are exceptionally high.

To understand why spatial context matters, it helps to distinguish between biomarkers that vary across a tumor and those that do not. The authors organize tissue-based biomarkers along a spectrum of spatial stability. HPV and p16 status, which define a biologically distinct and generally more treatable subtype of oropharyngeal cancer, tend to be uniform throughout the tumor, so a single biopsy usually captures them reliably. Certain genomic alterations behave similarly. But other clinically decisive biomarkers are far less cooperative. PD-L1 expression, the density and location of immune-cell infiltration, necrosis, and immune exclusion — the phenomenon where immune cells are kept at the tumor’s periphery — can all vary dramatically from one region of a tumor to another. A core needle sample taken from a well-perfused, immune-inflamed edge may paint a completely different picture than a hypoxic, necrotic core just a few millimeters away.

This is not a theoretical concern. The landmark multiregion sequencing work of Gerlinger and colleagues, published in the New England Journal of Medicine in 2012, demonstrated that intratumor heterogeneity and branched evolution are fundamental features of cancer biology. In HNSCC, the clinical consequences are concrete. When a tumor board decides whether a patient with HPV-positive oropharyngeal cancer can safely receive reduced-intensity treatment — so-called de-escalation — the decision often hinges on biomarkers measured in a single sample. If that sample happens to come from an unrepresentative region, the de-escalation decision rests on sand. Similarly, eligibility for immunotherapy frequently depends on PD-L1 scoring, and the review points out that localized sampling may incompletely capture heterogeneity in exactly these selected clinical contexts.

Radiomics offers a fundamentally different vantage point. The technique extracts hundreds of quantitative features from routinely acquired medical images — computed tomography, magnetic resonance imaging, and FDG-PET — converting the visual texture, shape, and intensity patterns of a tumor into high-dimensional data. First articulated in a widely cited 2014 Nature Communications paper by Aerts and colleagues, the premise is that image features reflect underlying biological processes: coarse heterogeneity on CT may mirror necrosis and hypoxia; specific texture patterns on diffusion-weighted MRI, measured through the apparent diffusion coefficient, correlate with cell density and oxygenation; FDG-PET textural features track metabolic activity and its spatial distribution. Because imaging captures the entire tumor volume — and often the surrounding peritumoral tissue — it sees what the biopsy cannot.

The review synthesizes a substantial body of evidence linking radiomic features to the biological processes that matter most in HNSCC. Tumor hypoxia is a classic example: it has been known since Brizel and colleagues’ 1997 work to worsen prognosis in head and neck cancer, largely because oxygen-deprived cells resist radiation. Hypoxia is also profoundly spatial — it develops in regions distant from functional blood vessels — which makes it a natural target for imaging. Studies have shown that textural features of hypoxia PET predict survival during chemoradiotherapy, and recent work on apparent diffusion coefficient MRI suggests it can act as a predictive biomarker for hypoxia, treatment de-escalation, and recurrence in HPV-associated oropharyngeal cancer. On the immune side, radiomic signatures have been associated with PD-L1 expression, CD8-positive T-cell infiltration, T cell-inflamed gene expression profiles, and even gamma-delta T-cell abundance, while radiogenomic analyses have connected imaging heterogeneity to somatic mutations in genes such as TP53, FAT1, and KMT2D.

What sets the Keio review apart from much of the radiomics literature is its insistence on mapping these features to specific clinical decision points rather than treating them as generic prognostic scores. The authors walk through three major scenarios. The first is larynx preservation, where the choice between organ-sparing chemoradiotherapy — a strategy established by the landmark 2003 trial of concurrent chemotherapy and radiotherapy — and primary surgery depends on predicting response. Radiomic features reflecting hypoxia and necrosis could flag tumors whose biopsy-based profiles look favorable but whose whole-tumor biology suggests resistance. The second is immunotherapy stratification, where combining a PD-L1 score from a biopsy with imaging evidence of immune exclusion or spatially restricted immune activity could refine patient selection for checkpoint inhibitors such as pembrolizumab and nivolumab, agents whose benefit in recurrent or metastatic HNSCC was established in the KEYNOTE-048 and CheckMate 141 trials. The third is recurrence assessment, where peritumoral radiomic signatures and subregion-based models have shown promise in predicting locoregional failure after chemoradiotherapy and even the site of recurrence after reirradiation.

The review also embraces a dynamic dimension of imaging that tissue sampling cannot match: delta radiomics, the analysis of how radiomic features change over time. Serial multiparametric MRI and FDG-PET acquired during the course of radiation therapy have been shown to predict treatment response, and FDG-PET can identify pathological response early during neoadjuvant immune checkpoint blockade. In principle, a clinician could watch the spatial biology of a tumor evolve week by week during treatment and adjust course accordingly — something no biopsy workflow could realistically achieve. Combined with deep-learning segmentation tools that automate tumor delineation on CT and MRI, and with harmonization methods such as ComBat that correct for differences between scanners and institutions, the technical pipeline for deploying these approaches at scale is maturing rapidly.

The authors are careful about the limits of the evidence. Nearly all of the studies they synthesize are retrospective, and the field has long struggled with reproducibility: radiomic features can be sensitive to image acquisition parameters, reconstruction algorithms, and segmentation choices, which is why standardization initiatives such as the Image Biomarker Standardisation Initiative have become essential infrastructure. Multi-institutional modeling studies have begun to evaluate how well radiomics and deep-learning models generalize beyond their development cohorts, and systematic reviews of machine-learning models for predicting radiation toxicity have taken a sober look at actual versus claimed performance. The review’s framing — radiomics as context rather than standalone predictor — is partly a response to this uncertainty. A model that claims to replace PD-L1 testing must clear an enormous evidentiary bar; a tool that flags when a biopsy result deserves a second look needs to be merely reliable enough to prompt caution.

That pragmatic framing may prove to be the review’s most lasting contribution. Rather than promising a revolution, Kono and colleagues describe an incremental integration: radiomics as a layer of spatial intelligence woven into existing biomarker-guided workflows, identifying the patients in which hypoxia, necrosis, stromal architecture, or immune exclusion make a small sample untrustworthy. For a disease where the difference between de-escalation and full-dose treatment can mean preserving a voice, and where immunotherapy decisions hinge on biomarkers that flicker across the tumor landscape, that kind of whole-tumor perspective could change how multidisciplinary teams weigh the evidence in front of them. The next step, the authors imply, is prospective validation — testing whether spatially informed caution actually improves outcomes. If it does, the humble biopsy may finally gain the companion it has always needed: a map of everything it cannot see.

Subject of Research: Use of radiomic imaging features as spatial context for interpreting biopsy biomarkers in head and neck squamous cell carcinoma treatment decisions

Article Title: Radiomics as a spatial context for treatment decision-making in head and neck cancer

Article References: Kono, T., Kasahara, K., Ogawa, R., & Ozawa, H. (2026). Radiomics as a spatial context for treatment decision-making in head and neck cancer. British Journal of Cancer. https://doi.org/10.1038/s41416-026-03600-0

Image Credits: AI Generated

DOI: 10.1038/s41416-026-03600-0

Keywords: radiomics, head and neck cancer, HNSCC, spatial heterogeneity, PD-L1, HPV, hypoxia, immunotherapy, biomarkers, de-escalation, medical imaging, tumor microenvironment