Towards real-time conformal palliative treatment of spine metastases: A deep learning approach for Hounsfield Unit recovery of cone beam CT images

Background

The extension of onboard cone-beam CT (CBCT) imaging for real-time treatment planning is constrained by limitations in image quality. Synthetic CT (sCT) generation using deep learning provides a potential solution to these limitations.

Purpose

This study was dedicated to creating a model capable of rapidly generating sCT images from CBCT scans, specifically for the entire spine. This work aims to be a step towards a CT simulation-free workflow by using onboard imaging for real-time palliative radiotherapy treatments for patients with spinal metastases.

Methods

Using CBCT and planning fan-beam CT images from 220 patients, we developed and validated a two-stage sCT generation model. The initial stage used a conditional generative adversarial network (GAN) to minimize streaking artifacts in CBCT images, using 7400 images for training and 1000 for validation. The second stage used a cycle-consistent GAN to produce sCT images, training on 14,700 images and validating on 500 images. The quality of the sCT images was evaluated quantitatively using a distinct dataset from 33 patients who received same-day palliative radiotherapy for spinal metastases.

Results

Our two-stage model generated high-quality sCT images from CBCT scans across the entire spine, significantly improving HU accuracy and dosimetric agreement with planning CT images. Mean Absolute Error was reduced from 225 $\pm$ 62 HU in CBCT to 86 $\pm$ 24 HU in sCT images, and Mean Error was improved from 178 $\pm$ 91 HU to −8 $\pm$ 20 HU. Dosimetric comparison for a subset of 20 patients indicated that the mean dose discrepancy for sCT-based calculations was lower than CBCT-based calculations by 4.5%, with the gamma (2 mm/2%) pass rate increasing by 34% on average.

Conclusions

This study demonstrates how a two-stage network facilitates CBCT-based sCT generation across the entire spine without prior CT knowledge, improving HU accuracy and potentially enabling highly-conformal palliative treatment planning for spinal metastases in real time.