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

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence

Background:

• Cone-beam computed tomography (CBCT) is essential in radiotherapy but poses a secondary cancer risk due to ionizing radiation, especially in pediatric patients.
• Deep learning super-resolution (SR) has improved image resolution but hasn't been applied to reduce CBCT radiation dose.

Purpose of the Study:

• To reduce CBCT imaging dose by employing an enhanced super-resolution generative adversarial network (ESRGAN).
• To restore image quality in low-dose CBCT using ESRGAN in both projection and image domains.

Main Methods:

• Trained two ESRGAN models on 2997 head and neck cancer CBCT scans: one in the projection domain (CBCT_SRpro) and one in the image domain (CBCT_SRimg).
• Evaluated SR CBCTs for image similarity, noise, spatial resolution, and registration accuracy against original high-dose CBCT (CBCT_HR).
• Conducted a visual Turing test to assess perceptual differences between original and SR CBCTs.

Main Results:

• Both projection and image domain SR improved low-dose CBCT quality; visual Turing tests showed minimal perceptual difference.
• CBCT_SRimg slightly outperformed CBCT_SRpro in the visual Turing test.
• SR methods significantly improved spatial resolution (CBCT_SRpro: 0.88 lp/mm, CBCT_SRimg: 0.95 lp/mm) compared to low-dose CBCT (CBCT_LR: 0.66 lp/mm), approaching high-dose levels (CBCT_HR: 1.01 lp/mm).
• Noise characteristics and registration accuracy for SR CBCTs were comparable to high-dose CBCT.

Conclusions:

• Deep learning super-resolution is a viable method for reducing CBCT dose in radiotherapy.
• Image domain SR processing generates higher quality images than projection domain processing for low-dose CBCT.
• This approach enables acquiring low-dose CBCTs while maintaining diagnostic image quality, crucial for patient safety.