Thyroid cancer and radioiodine: model first, learn only what is missing

How many sessions of radioactive iodine, how far apart, and at what activity? An intelligible model of three equations explores these questions before any patient is treated. Learning will be added only where the model is not enough.

Marie Fusella Giuntini and Dominique Barbolosi, with Cyril Voyant and David Taieb

The finding

The model and its simulator, RAIR-Sim, are the work of Marie Fusella Giuntini and Dominique Barbolosi, with Cyril Voyant and David Taieb (Scientific Reports, 2026). When a disease can be described by an intelligible model, that model comes first: it explains, it can be calibrated on real patients, and it lets one test protocols in silico. Artificial intelligence is not needed to do what the model already does well.

The question

Metastatic differentiated thyroid cancer is often treated with repeated sessions of radioactive iodine. Some tumours respond; others become refractory. Thyroglobulin (Tg), measured in the blood, follows the tumour burden. Which protocol, and for which patient profile?

The theory, in plain words

Three quantities evolve over the months:

• the activity of the administered iodine decreases over time, and each session adds a new dose to what remains;
• the tumour cells grow at a rate set by their doubling time T_d, and are killed in proportion to the activity present;
• thyroglobulin is produced by the tumour cells and cleared by the body.

The equations are solved session by session, so a whole protocol can be simulated in seconds. A short doubling time means a fast-growing tumour: that is what separates a refractory profile from a responding one.

Where it sits on the complexity ladder

Second rung: simulation from a model we understand. The next step is a hybrid digital twin: keep the model, and let a learning method (a frugal extreme learning machine, for example) estimate only what the model does not capture. This is mathematics- and medicine-informed learning, not a black box replacing the model. No learning added where the model meets the need; learning, in hybrid form, where it helps.

The example in figures

With the public demonstrator parameters, 3.7 GBq per session every 3 months:

• Refractory profile (T_d = 9.8 months): thyroglobulin reaches about 1,386 ng/mL at 79 months after one session, and stays near 12 ng/mL after eight.
• Responding profile (T_d = 66.6 months): about 14 ng/mL after one session, and 0.13 ng/mL after eight.

These are model outputs for two archetypes, not predictions for a patient.

What it brings, and its limits

• Brings: an intelligible way to compare protocols before treating; a biomarker reading (the doubling time) that separates profiles; open code that anyone can run (MATLAB).
• Limits: two archetypes and population-calibrated parameters; it is not a clinical decision tool and needs prospective validation. The learning part of the hybrid twin is work in progress, separate from the published results.

Go further

• Article: M. Fusella Giuntini, C. Voyant, D. Taieb, D. Barbolosi, A computational framework for optimizing radioiodine therapy protocols in metastatic thyroid cancer, Scientific Reports 16, 27179 (2026), doi:10.1038/s41598-026-56267-1
• Full equations, results of the article and runnable code: https://cyrilvoyant.github.io/methods/thyroid-radioiodine.html
• Code (RAIR-Sim): github.com/MarieFG49/CTD_SIMU_RAI (MIT) · documented fork archived at doi:10.5281/zenodo.21496198
• Review: M. Fusella Giuntini, C. Voyant, D. Taieb, D. Barbolosi, Computational modeling and trends in RAI therapy for thyroid cancer, Biomedical Journal of Scientific & Technical Research 62(4) (2025), doi:10.26717/BJSTR.2025.62.009789
• For everyone (in French): "IA et cancer de la thyroïde : demain, la fin des traitements standardisés ?", The Conversation, May 2026
• The approach behind this work: Understand before you compute