ESTRO Course 2026 - Quantitative methods in Radiation Oncology: "Models in the Clinic"
Year: 2026
Type: article
Abstract: Delivered 2026-04-22, Lisbon, Portugal, Faculdade de Ciências da Universidade de Lisboa, Campo Grande. PrecisThis lecture, delivered for the ESTRO School's Quantitative Methods in Radiation Oncology course, examines the translational arc of predictive models in radiation oncology—from in silico development through in vivo clinical deployment—using head and neck cancer as the exemplar disease site. The talk opens by situating current practice within the consensus-based QUANTEC/HYTEC framework, noting that these normal tissue complication probability models, while clinically serviceable, predate modern IMRT and conceal substantial inter-institutional variability in dose-volume patterns, as demonstrated by the LAMBDA Consortium.Decision support tools—ORATOR2 and TORS vs. chemoradiation selection models for HPV-associated oropharyngeal cancer—illustrate the clinical stakes of miscalibrated risk quantification, given that humans perform near chance at pre-therapeutic extranodal extension detection and that deep learning consistently outperforms expert radiologists (13–23% AUC gain in E3311 validation). Longitudinal MDADI trajectories across ECOG 3311 arms further underscore how poorly we quantify post-treatment dysphagia risk, compounding the stakes of treatment selection. Despite this, clinical uptake of predictive tools remains limited by trust deficits. A parallel survey of auto-segmentation benchmarks—from the 2010 MICCAI parotid challenge through contemporary multiparametric MRI-based GTV delineation—documents persistent algorithm-observer disagreement and the limits of Dice-based evaluation when "better but different" models deviate from human reference standards, as with noise-injected brachial plexus avoidance.The central thesis is that clinical progress demands uncertainty quantification: distinguishing aleatoric (data) from epistemic (model) uncertainty via Bayesian neural networks, ensembles, or test-time augmentation, thereby enabling direct ISO 14971:2019 risk estimation at the patient level. Exemplar applications include oropharyngeal GTV segmentation with voxel-level uncertainty maps, web-based individualized survival prediction, longitudinal symptom burden forecasting (THALIS; LSTM-Cox frameworks), and digital twin visualization (DITTO).The closing argument frames clinical implementation as a socio-technical undertaking governed by the TRIPOD-AI, STARD-AI, DECIDE-AI, CONSORT-AI, and SPIRIT-AI reporting ecosystem, use-case-specific acceptance testing, and error-mode vigilance—demanding multidisciplinary exploration, solution-design, and implementation teams rather than checklists alone.
Source: Figshare
Author: Fuller, Clifton D.
Subfield: Radiation
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7155170634
FWCI:
Cites:
Cited by:
Related to: