The potential of Artificial Intelligence (AI) in radiation oncology seems tremendous with a vast array of
applications, ranging from refinement in diagnostic and disease delineation and staging to optimisation
of the clinical workflow, routine integration of multi-parametric and multi-dimensional imaging,
treatment response prediction, treatment guidance etc. Several types of gains are expected out of
these: shorten patients’ workflows while improving their prognosis, integrate in the routine high-
precision adaptive radiation therapy approaches, reduce treatment-associated costs, etc. Importantly,
AI applied to cancer care should ultimately lead to treatment harmonization, giving rise to procedures
that are less operator-dependent and consequently, progressively, to the reduction of care disparities
across centers.
However, such promises cannot be taken for granted. The conventional development path requires
methodologically-suited quantitative and qualitative evaluations, including important validation steps
in independent cohorts, that are key for envisioning further large-scale deployment. For now, AI-based
tools suffer from a lack of adhesion of the community overall, intrinsically linked to an insufficient
diffusion of novel AI technologies.
Confusion remains between Technology Readiness Level (TRL), the actual availability of a tool and its
clinical validation. As such, every novel AI-based techniques can, and should be evaluated compared
to the techniques of reference. This evaluation will be of key importance for several aspects:
convincing care providers and institutions of the value of a given technology with explainable metrics
(improved patients’ quality of life and clinical outcomes, decreased need of dedicated resources,
reduced time needed per patients, improved accuracy for the definition of patient staging and
prognosis etc.). Beside these relatively obvious and straightforward objectives, the evaluation of novel
AI has to satisfy the recently updated E.U regulation.
Another key aspect of the evaluation of AI tools is to enlighten the equipment choice for end-users,
often lost in the massive diversity of emerging tools, culminating in solicitation for constant
updates/upgrades. The move for AI-based technologies not only comes at a financial cost but also with
a change in practice; the investment should come with a tangible benefit.
We will present our experience with AI-based technology implementation in our institution, focused
on three examples: auto-contouring, pseudo-CT and immunological optimization of radiotherapy. We
will address the critical steps for a straightforward validation assessment, the strategy chosen for the
selection of metrics and their anticipated impact on clinical practice. We will also comment on the
needs for a robust medico-economics assessment – that may also include the carbon performances
and environmental concerns, if any – to provide the payers the arguments for any additional
investment.