A modality conversion approach to MV-DRs and KV-DRRs registration using information bottlenecked conditional...

Area of Science:

• Medical Imaging
• Radiation Oncology
• Artificial Intelligence

Background:

• Electronic portal imaging devices (EPIDs) are crucial for patient positioning in radiation therapy but produce low-contrast images (MV-DRs).
• Registering MV-DRs with kilovoltage digital reconstructed radiographs (KV-DRRs) is challenging due to image ambiguity and appearance variations.
• Accurate multimodal registration is vital for precise radiotherapy delivery.

Purpose of the Study:

• To develop a novel modality conversion approach for synthesizing KV images from MV-DRs.
• To improve the accuracy of multimodal registration between MV-DRs and KV-DRRs.
• To address challenges in deep-learning-based image synthesis for medical applications.

Main Methods:

• A conditional generative adversarial network with information bottleneck extension (IB-cGAN) was developed for synthesizing KV images.
• The IB-cGAN takes MV-DRs and nonaligned KV-DRRs as input to generate synthesized KV images.
• The model uses adversarial loss for semantic-level supervision and an information bottleneck to constrain information from nonaligned KV-DRRs.

Main Results:

• Trained on 2698 patient scans and tested on 208, the model synthesized realistic KV images.
• Qualitative results showed synthesized images enable visual registration.
• Quantitative analysis demonstrated a 22.37% improvement in registration accuracy compared to non-modality conversion methods.

Conclusions:

• The proposed modality conversion facilitates MV-KV registration for clinicians and algorithms.
• This approach can enhance image-guided radiation therapy, particularly in regions utilizing affordable EPIDs.
• It offers a pathway to improve radiotherapy accuracy and accessibility globally.