A patient-independent CT intensity matching method using conditional generative adversarial networks (cGAN) for...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Radiation Oncology

Background:

• Convolutional neural network (CNN)-based tumor localization using single X-ray projections is sensitive to intensity discrepancies between digitally reconstructed radiographs (DRRs) and measured X-ray projections.
• Previous intensity matching methods required patient-specific 3D-cone-beam CT (3D-CBCT) scans, which are inefficient and hinder clinical implementation.

Purpose of the Study:

• To develop and validate a patient-independent intensity matching method for 3D computed tomography (3D-CT) to improve tumor localization accuracy.
• To leverage a conditional generative adversarial network (cGAN) to generate synthetic 3D-CBCT data, bypassing the need for actual CBCT scans.

Main Methods:

• A 3D cGAN was trained on prior patient data to learn the mapping from 3D-CT to 3D-CBCT.
• The trained cGAN was applied to new patients to generate synthetic 3D-CBCT, from which synthetic DRRs were created.
• The synthetic DRRs were used in a CNN-based tumor localization framework, and results were compared to actual measurements.

Main Results:

• The synthetic 3D-CBCT showed a median intensity difference of ≤10 HU compared to real 3D-CBCT across all patients.
• The relative error between synthetic DRRs and measured X-ray projections was <4.8% ± 2.0%.
• Tumor localization errors for patients with visible tumors were <1.7 mm (superior-inferior) and <0.9 mm (lateral).

Conclusions:

• The proposed patient-independent CT intensity matching method, based on cGANs, enables accurate tumor localization.
• This method significantly improves clinical workflow efficiency by eliminating the requirement for pre-treatment CBCT scans for each patient.
• The approach offers a promising solution for robust and efficient tumor localization in radiation therapy.