ESTRO 2024 - Abstract Book

S3872

Physics - Image acquisition and processing

ESTRO 2024

MR-only workflow in the brain: a DL algorithm to generate synthetic CT starting from 0.35T MRI

Luca Vellini 1 , Jacopo Lenkowicz 2 , Claudio Votta 3 , Sebastiano Menna 1 , FlavioVincenzo Quaranta 1 , Elisa Pilloni 1 , Michele Aquilano 4 , Andrea D'Aviero 4 , Martina Iezzi 4 , Francesco Preziosi 4 , Alessia Re 4 , Althea Boschetti 5 , Danila Piccari 5 , Antonio Piras 6 , Carmela Di Dio 4 , Francesco Catucci 4 , Gian Carlo Mattiucci 4 , Davide Cusumano 1 1 Mater Olbia Hospital, Physics, Olbia, Italy. 2 Policlinico Gemelli, Physics, Roma, Italy. 3 Policlinico Gemelli, TSRM, Roma, Italy. 4 Mater Olbia Hospital, Radiotherapist, Olbia, Italy. 5 Mater Olbia Hospital, TSRM, Olbia, Italy. 6 Villa Santa Teresa, Radiotherapist, Palermo, Italy

Purpose/Objective:

MR-only workflow represents the last frontier of Magnetic Resonance guided Radiotherapy (MRgRT), achievable thanks to the integration of Artificial Intelligence (AI) in clinical workflow. Removing computed tomography (CT) simulation from the clinical workflow can be obtained using novel Deep Learning (DL) algorithms, able to generate synthetic CT (sCT) from on-board MR images in time suitable for online adaptation (less than 2 minutes). The aim of this study is to propose a new DL algorithm able to generate sCT from low field MR images in the brain, which represents an unexplored site in the field of low field MRI.

Material/Methods:

A total of 40 patients were enrolled for this study and divided in training (20), validation (10) and test (10) sets. Patients were treated using a low field MR-Linac system, acquiring a 0.35 T T2*/T1 MRI and a CT images during treatment simulation. A multistep pre-processing phase was applied to all images, consisting of bias field correction, histogram equalisation and spatial resampling. Image selection was performed slice by slice before moving to the training phase. A conditional Generative Adversarial Network (cGAN) was trained on axial paired images, which were converted in PNG format to make them compliant for network architecture. Image pre-processing was performed using Rstudio, network training and validation using Python and Tensorflow. Hyperparameters tuning was performed during the training phase to find the optimal values in terms of learning rate, batch size and lambda. All the training phases were run for 200 epochs. Once trained the neural network, the test cases generated sCT underwent image accuracy analysis, which was estimated calculating the mean absolute error (MAE) and the mean error (ME) in terms of Hounsfield Units (HU) between synthetic and original CT. Two Intensity Modulated Radiation Therapy (IMRT) treatment plans were calculated for each patient, both considering the simulation MRI as reference image: the first plan was calculated using the original CT for Electron Density (ED) map, the second one using the sCT. Plans considered in the analysis included a hypofractionation regimens, with total dose ranging from 25-27 Gy in 3 to 5 fractions. Dose calculation was performed considering the presence of magnetic field, using a Montecarlo algorithm and a dose grid size of 1.5 mm3. A comparison between the two IMRT plans was performed in terms of gamma analysis and Dose Volume Histogram (DVH) differences. Four PTV parameters (V95%, D2%, D98% and D50%) and three parameters for spinal cord (D2%, D98% and D50%) were compared evaluating plans DVH.

Results: