Deep learning based super-resolution for CBCT dose reduction in radiotherapy

Cone-beam computed tomography (CBCT) is a crucial daily imaging modality in image-guided and adaptive radiotherapy. However, the use of ionizing radiation in CBCT imaging increases the risk of secondary cancers, which is particularly concerning for pediatric patients. Deep learning super-resolution has shown promising results in enhancing the resolution of photographic and medical images but has not yet been explored in the context of CBCT dose reduction.

To facilitate CBCT imaging dose reduction, we propose using an enhanced super-resolution generative adversarial network (ESRGAN) in both the projection and image domains to restore the image quality of low-dose CBCT.

An extensive projection database, containing 2997 CBCT scans from head and neck cancer patients, was used to train two different ESRGAN models to generate super-resolution CBCTs. One model operated in the projection domain, using pairs of simulated low-resolution (low-dose) and original high-resolution (high-dose) projections and yielded CBCT_SRpro. The other model operated in the image domain, using pairs of axial slices from reconstructed low-resolution and high-resolution CBCTs (CBCT_LR and CBCT_HR) and resulted in CBCT_SRimg. Super-resolution CBCTs were evaluated in terms of image similarity (MAE, ME, PSNR, and SSIM), noise characteristics, spatial resolution, and registration accuracy, using the original CBCT as a reference. To test the perceptual difference between the original and super-resolution CBCT, we performed a visual Turing test.

Visually, both super-resolution approaches in the projection and image domains improved the image quality of low-dose CBCTs. This was confirmed by the visual Turing test, that showed low accuracy, sensitivity, and specificity, indicating almost no perceptual difference between CBCT_HR and the super-resolution CBCTs. CBCT_SRimg (accuracy: 0.55, sensitivity: 0.59, specificity: 0.50) performed slightly better than CBCT_SRpro (accuracy: 0.59, sensitivity: 0.61, specificity: 0.57). Image similarity metrics were affected by varying noise levels and did not reflect the visual improvements, with MAE/ME/PSNR/SSIM values of 110.4 HU/2.9 HU/40.4 dB/0.82 for CBCT_LR, 136.6 HU/-0.4 HU/38.6 dB/0.77 for CBCT_SRpro, and 128.2 HU/1.9 HU/39.0 dB/0.80 for CBCT_SRimg. In terms of spatial resolution, quantified by calculating 10% levels of the task transfer function, both CBCT_SRpro and CBCT_SRimg outperformed CBCT_LR and nearly matched the reference CBCT_HR (CBCT_LR: 0.66 lp/mm, CBCT_SRpro: 0.88 lp/mm, CBCT_SRimg: 0.95 lp/mm, CBCT_HR: 1.01 lp/mm). Noise characteristics of CBCT_SRimg and CBCT_SRpro were comparable to the reference CBCT_HR. Registration parameters showed negligible differences for all CBCTs (CBCT_LR, CBCT_SRpro, CBCT_SRimg), with average absolute differences in registration parameters being below 0.4° for rotations and below 0.06 mm for translations (CBCT_HR as reference).

This study demonstrates that deep learning can be a valuable tool for CBCT dose reduction in CBCT-guided radiotherapy by acquiring low-dose CBCTs and restoring the image quality using deep learning super-resolution. The results suggest that higher quality images can be generated when super-resolution is performed in the image domain compared to the projection domain.