Minimum imaging dose for deep learning-based pelvic synthetic computed tomography generation from cone beam images

Abstract

Background and purpose

Daily cone-beam computed tomography (CBCT) in image-guided radiotherapy administers radiation exposure and subjects patients to secondary cancer risk. Reducing imaging dose remains challenging as image quality deteriorates. We investigated three imaging dose levels by reducing projections and correcting images using two deep learning algorithms, aiming at identifying the lowest achievable imaging dose.

Materials and methods

CBCTs were reconstructed with 100%, 25%, 15% and 10% projections. Models were trained (30), validated (3) and tested (8) with prostate cancer patient datasets. We optimized and compared the performance of 1) a cycle generative adversarial network (cycleGAN) with residual connection and 2) a contrastive unpaired translation network (CUT) to generate synthetic computed tomography (sCT) from reduced imaging dose CBCTs. Volumetric modulated arc therapy plans were optimized on a reference intensity-corrected full dose CBCT_cor and recalculated on sCTs. Hounsfield unit (HU) and positioning accuracy were evaluated. Bladder and rectum were manually delineated to determine anatomical fidelity.

Results

All sCTs achieved average mean absolute mean absolute error/structural similarity index measure/peak signal-to-noise ratio of $⩽$59HU/$⩾$0.94/$⩾$33 dB. All dose-volume histogram parameter differences were within 2 Gy or 2$\%$. Positioning differences were $⩽$0.30 mm or 0.30°. cycleGAN with Dice similarity coefficients (DSC) for bladder/rectum of $⩾$0.85/$⩾$0.81 performed better than CUT ($⩾$0.83/$⩾$0.76). A significantly lower DSC accuracy was observed for 15$\%$ and 10$\%$ sCTs. cycleGAN performed better than CUT for contouring, however both yielded comparable outcomes in other evaluations.

Conclusion

sCTs based on different CBCT doses using cycleGAN and CUT were investigated. Based on segmentation accuracy, 25$\%$ is the minimum imaging dose.