S2297
Physics - Machine learning and AI algorithms
ESTRO 2026
Digital Poster 2742 Improving treatment planning outcomes with AI tools in radiotherapy Noemi Cucurachi 1 , Paola Tabarelli De Fatis 2 , Chiara Pellegrini 2 , Giada Tognolini 3 , Giampietro Barai 1 , Daniela D'Ambrosio 2 1 Medical Physics, ASST Mantova, Mantova, Italy. 2 Medical Physics, ICS Maugeri, Pavia, Italy. 3 Medical Physics, University of Milan, Milan, Italy Purpose/Objective: Autoplanning systems standardize treatment planning and reduce working times. However results aren’t always optimal. This study aims to predict outcomes of commercial autoplanning system using artificial intelligence and implementing a new optimization
model accuracy, with MAE reducing from 42.0±9 HU to 35.9±6 HU. MRI pre-processing showed minimal impact (p=0.45 between filtered and raw images), while the choice of architecture was critical (Figure 2): AFNO achieved the lowest MAE (32.4±6 HU, p<0.01 with the other generator architectures). The best network (AFNO trained on raw MR images from 102 patients) showed high dose accuracy with dose deviations within 3% for PTV and 50cGy for the OARs indicators.
workflow for failures. Material/Methods:
119 prostate cancer patients were planned (60 Gy in 20 fractions) at ICS Maugeri of Pavia. Each patient was assigned a binary outcome reflecting the autoplanning success in reaching optimal target coverage, respecting OARs constraints. Six ANOVA anatomical features from 84 CT exams were used to train and evaluate machine learning models (MLMs) in MATLAB v.2023b; 20 patients were used for external testing. Three MLMs (linear SVM, subspace discriminant, and logistic regression)were selected based on AUC, accuracy, precision and recall. A home-made MATLAB GUI integrating the three MLMs through a bagging (majority voting) strategy was developed and validated on 15 patients. A new optimization workflow was created to improve CTV and PTV coverage in failed cases; T-tests were computed to evaluate target coverage differences between the two workflows. A convolutional neural network (CNN) was developed usingPyTorch library. The CNN was trained on 101 CT exams using transfer learning approach through the pre-trained MedicalNet repository. All images were standardized, and data augmentation techniques were applied.Its performance was assessed through accuracy, AUC, precision and recall, and activation maps were generated to evaluate the network's behaviour.
Conclusion: AFNO markedly improves thoracic sCT generation, outperforming traditional architectures in both image quality and dosimetric accuracy. Its integration into MRI-only workflows shows strong clinical potential. Nonetheless, further validation across larger, multicentre cohorts is essential to confirm robustness widespread adoption in thoracic MRIgRT. Keywords: Synthetic CT, Lung, GAN
Made with FlippingBook - Share PDF online