ESTRO 2025 - Abstract Book

S3066

Physics - Image acquisition and processing

ESTRO 2025

acquisition time, taking over four minutes. We propose employing the quickDWI 3 deep learning (DL) algorithm 4 to accelerate and denoise diffusion-weighted (DW) MRI on an MR-Linac.

Material/Methods: We collected data from 12 prostate cancer patients who consented to participation in the HERMES trial (REC 20/LO/1162), which involved two to four scans at various treatment stages on a 1.5T Unity MR-Linac (Elekta AB, Stockholm). Model training used data from eight patients (20 scans), with two patients each for validation (7 scans) and testing (8 scans). The acquired images (224×224 pixels) covered 15 slices with b-values (averages) of 0(6), 30(6), 150(6), 500 (14) s/mm² and 3 DW directions. Individual images for each average and DW direction were exported for training. Six distinct models were trained: • All directions: AD1, AD2, AD3, where one, two or three averages from each direction were used as input data for each b-value; • Single direction: SD1, SD2 and SD3, where only data from one direction was used. A U-net architecture was trained with mean absolute error (MAE) loss, learning rate 1e-4 and batch size 30 for 50 epochs. We assessed MAE and root mean square error (RMSE) of the ADC maps and compared ADC distributions within the prostate.

Results: