ESTRO 2024 - Abstract Book

S3930

Physics - Image acquisition and processing

ESTRO 2024

2547

Proffered Paper

Projection-based deep learning super-resolution for CBCT dose reduction

Adrian Thummerer 1 , Jan Hofmaier 1 , Claus Belka 1,2,3 , Guillaume Landry 1 , Christopher Kurz 1

1 LMU University Hospital, LMU Munich, Department of Radiation Oncology, Munich, Germany. 2 German Cancer Consortium (DKTK), partner site Munich, a partnership between DKFZ and LMU University Hospital Munich Germany, Munich, Germany. 3 Bavarian Cancer Research Center, (BZKF), Munich, Germany

Purpose/Objective:

CBCT imaging has become an essential imaging modality in modern image-guided radiotherapy, as it provides the necessary volumetric information for pre-treatment patient alignment and has recently enabled daily adaptive radiotherapy workflows [1]. However, the large number of CBCTs acquired throughout the fractionated treatment results in a considerable imaging dose to the patient, not only in regions that receive significant treatment dose but also in healthy tissues outside the treatment field [2].

In this study we propose the use of a generative adversarial network (GAN) to perform super-resolution in the projection domain to enable the acquisition of low dose CBCTs with negligible loss in image quality.

Material/Methods:

An extensive dataset of 592776 CBCT projections (XVI, Elekta, Sweden) from 2997 head and neck cancer patients was used to train and evaluate an enhanced super-resolution GAN (ESRGAN) [3] to up-sample low-dose, low-resolution CBCT projections. Low dose projections were simulated by a noise preserving down-sampling method that increases the pixel spacing from 0.8 mm to 1.6 mm and reduces the image dimensions from 504x504 to 252x252. In theory, this down-sampling approach simulates a fourfold dose reduction. In addition, CBCT scans of a water phantom were acquired with the same clinical head and neck imaging protocol (10mA, 10ms and 100kV) to systematically evaluate the resolution and noise characteristics of super-resolution CBCTs (not included in the training dataset). After training ESRGAN with 495070 pairs of low- and high-resolution projections, evaluation was performed on 250 test patients (49484 projections). CBCT HR was reconstructed from the original CBCT projections, CBCT LR from the simulated low-dose, low-resolution projections, CBCTSR from the deep learning up-sampled projections and CBCT BI from conventionally up-sampled projections using bilinear interpolation. All reconstructions were performed on a 270x270x264-pixel grid with isotropic pixel spacing of 1mm with the FDK algorithm. The reconstructed CBCTs were compared in terms of image similarity using mean absolute error (MAE), mean error (ME) and peak signal-to-noise ratio (PSNR) using CBCT HR as ground truth. The similarity of bone structures was assessed by generating a bone mask on each CBCT and calculating the Dice similarity coefficient (DSC) with CBCT HR as reference image. To mimic the patient alignment procedure a rigid registration between each CBCT and the planning CTs was performed, and the resulting registration parameters were compared. Resolution and noise