2026 ESTRO Artificial Intelligence in Radiotherapy Clinical Practice: "From voxels to predictions:Radiomics and dosiomics"

From Voxels to Predictions: Radiomics and Dosiomics — PrécisRadiomics in radiation oncology sits at something of a precipice. A decade-plus after Aerts et al. (2014), the reasonable observer could still fairly ask whether the field is “voodoo science,” overhyped, or merely a research curiosity. This talk takes that question seriously rather than defensively.The first half is a (candid) retrospective. Head and neck radiomics has shown reproducible signal for classification tasks (HPV/TP53 status, tumor phenotype), and incremental gains in locoregional control models, but essentially nothing for overall survival. The 2016 MICCAI/Kaggle OPC challenges are instructive, insofar as the local-control winner discarded the imaging altogether and used staging and demographics. Preclinical GADD34 work shows radiomic features can in fact track genomically-driven radiosensitivity; whether that adds clinical actionability is a separate matter. The subsequent methodologic literature (scanner variability, volume-confounding, partition leakage, the Radiomics Quality Score) explains much of the non-reproducibility: radiomics is simply hard to do well, and acquisition decisions made upstream of any model constrain what can be learned downstream.The normal-tissue examples are, I’d argue, where the approach has earned its keep, i.e., CT-based osteoradionecrosis prediction, and van Dijk’s parotid work showing that PET/MR intensity and delta-surface features improve xerostomia NTCP models, with external validation (UMCG/MDACC) to boot. Notably, the same series demonstrates that most texture delta-features were dose surrogates driven by one or two outliers, which is precisely the kind of negative finding the field needs.Dosiomics is then introduced as the extension of these principles to 3D dose, motivated by Bellman’s curse of dimensionality: collapsing a dose distribution to a DVH, then to a mean, then to “dry mouth, Y/N” is an exercise in serial information loss. Subvolume, beam-path, and 3D deep-learning NTCP examples (dysphagia, temporal lobe injury, pneumonitis) illustrate the alternative.The closing argument is that the way forward requires both better models (probabilistic, causal, mechanistic; FAIR data) and better testing (physics-led trials, cross-disciplinary engagement, and a willingness to publish null results). The tools are maturing; so must we.