Author: Wouter Crijns, Frederik Maes, Loes Vandenbroucke, Liesbeth Vandewinckele 👨🔬
Affiliation: Department of Oncology, Laboratory of Experimental Radiotherapy, KU Leuven; Department of Radiation Oncology, UZ Leuven, Department ESAT/PSI, KU Leuven; Medical Imaging Research Center, UZ Leuven, Department of Oncology, Laboratory of Experimental Radiotherapy, KU Leuven 🌍
Purpose: To explore intentional deep overfit learning (IDOL) to exploit the initial treatment plan to predict an adaptive radiotherapy plan.
Methods: A conditional generative adversarial network is used to predict fluence maps using patient anatomical information, summarized using projections of the OARs and PTV [1]. Following the IDOL framework [2][3], a general model was trained on the initial and mid-treatment (adaptive) data of 111 IMRT-treated lung cancer patients, and overfit to initial, patient-specific data excluded from training. The patient-specific model is used for adaptive fluence prediction. Fluence maps of the general and the patient-specific model were compared to the adaptive clinical plan using an independent test dataset of 18 patients. Differences in fluence metrics are analyzed according to the Wilcoxon signed-rank test, and DVHs were calculated in the treatment planning system.
Results: Comparing predicted fluence maps from the general and the patient-specific model showed improved fluence metrics (structural similarity index 0.977 to 0.981 [p<0.05], gamma pass rate 0.969 to 0.975 [p<0.05] and mean absolute error 0.0033 to 0.0027 [p<0.05]). DVH comparison showed improved, albeit inferior, PTV coverage in the patient-specific model with D95%= 54.6, 55.8 and 61.8 Gy; Dmean= 64.2, 63.8 and 64.2 Gy for general, patient-specific, and clinical reference, respectively. OAR doses were similar across models: Dmax= 23.7, 22.9 and 23.6 Gy for the mediastinal envelope and Dmean= 12.9, 12.4 and 12.4 Gy for the lungs, for general, patient-specific and clinical reference, respectively.
Conclusion: Exploiting the initial treatment plan improves fluence prediction for adaptive plans, allowing for a patient-specific adaptive radiotherapy approach. PTV coverage is not yet sufficiently similar to most published models [4][5][6]. The patient-specific model supports a longitudinal approaches to adaptive radiotherapy, embedding patient-specific data over time.
References:
[1]X.Li etal.,PhysMedBiol,vol.65,no.17,p.175014,Aug.2020
[2]J.Chun etal.,MedPhys,vol. 49,no.1,pp.488–496,Jan.2022
[3]A.Maniscalco etal.,Me Phys,vol.50,no.9,pp.5354–5363,Sep.2023
[4]W.Wang etal.,AdvRadiatOncol,vol.6,no.4,p.100672,Jul.2021
[5]H.Lee etal.,ScientificReports20199:1,vol.9,no.1,pp.1–11,Oct.2019
[6]Y.Li etal.RadiationOncology,vol.18,no.1,pp.1–12,Dec.2023