Session Item

In RayStation v11A® TPS, 3 methods are available to generate treatment plans: A Multi-Criteria Optimization (MCO) and two Artificial Intelligences (AI), a Machine Learning (ML) and a Deep Learning (DL) algorithm. The goal is to evaluate these advanced methods to automatize our clinical routine.

The ML is based on Atlas of Random Forest and the DL on a Fully Convolutional Neural Network. They generate a predicted dose distribution. The AI methods were both trained on local treatment plans database, then the obtained models were adjusted in a dedicated RaySearch module, RayLearner.

The MCO algorithm generate treatment plans (from users constraints) were no optimization goals can be improved without deteriorating another: Optimal Pareto Plans. The chosen predicted dose distribution calculated from theorical fluence.

Then, the three methods use the same Mimicking algorithm which transforms the predicted dose distribution into a deliverable treatment plan.

Methods have been evaluated on prostate cancer treated at 78, 76 or 66 Gy, delivered with VMAT arcs of photons (6 or 10MV) on Novalis TX and Halcyon VARIAN®. We developed scripts to use MCO automatically and the AI algorithms were trained with 100 of our 78Gy treatment plans and they were adjusted in RayLearner to be used at 76 and 66Gy.

Plans were clinically accepted when they fullfil the RECORAD recommendations and pass the Patient Quality Assurance Criteria: Gamma index > 1 for 95% of points in 3% 2mm, measured with ARCHECK (and/or EPID SunCheck system (SunNuclear)

We compared MCO, ML and DL plans to manual Standard Optimized plans (SO) on 50 treatment plans with dosimetric Index: Conformity (C), Homogeneity (H) and Dose Gradient (DG). And the plan complexity were compared with the Modulation Complexity Score (MCS). We also evaluate the plan generation duration: active time (required by the user) and passive time (without the user). 

SO plans were clinically accepted with index : C≈ 0.84, H≈0.07, DG ≈3.11, MSC ≈ 0.35 and plans were generated in an average time of 180 active minutes.

80% of ML plans were accepted after 16 passive minutes and 85% of DL plans were accepted after 11 passive minutes. 100 % of others AI plans were accepted after additionnal optimization (≈ 5 active minutes). AI dose distributions were at least equivalent and more complex  than SO: C≈ 0.86, H≈0.08, DG ≈3.02, MSC ≈ 0.31

100% of MCO plans were generated in 30 passive minutes and were after additionnal standard optimization (≈ 30 active minutes). MCO plans had the best and most complex dose distributions (C≈ 0.87, H≈0.06, DG ≈2.88, MSC ≈ 0.2)

Automated methods gives results more homogeneous and very close (DL / ML) or event better (MCO) than SO. Futhermore, calculation were 2(MCO) to 16 (DL) times faster. Finally, MCO seems to be more adapted to complex and unusual clinical cases and AI can be used on classical clinical cases which allows to consider adaptative radiotherapy.