CrossModalNet: exploiting quality preoperative images for multimodal image registration.

Area of Science:

• Medical Imaging
• Radiotherapy
• Computer Vision

Background:

• Image-guided radiotherapy (IGRT) relies on accurate image registration.
• Intraoperative digital radiography (DR) images often suffer from low quality (blur, noise, low contrast).
• This quality degradation challenges automatic registration algorithms, particularly for DR-DRR multimodal registration.

Purpose of the Study:

• To develop a novel Convolutional Neural Network (CNN)-based method, CrossModalNet, to improve DR-DRR registration accuracy.
• To leverage high-quality preoperative digitally reconstructed radiographs (DRR) to compensate for limitations in intraoperative digital radiography (DR) images.
• To enhance the robustness of automatic registration algorithms in IGRT.

Main Methods:

• Proposed CrossModalNet, a CNN-based approach for DR-DRR registration.
• Incorporated CrossModal Attention and Refine Modules to exploit multiscale crossmodal features and interactions.
• Implemented a two-part method: DR-DRR contour prediction followed by contour-based rigid registration using mutual information.
• Trained on 2486 patient scans and tested on 170 scans.

Main Results:

• CrossModalNet demonstrated superior performance compared to classic and state-of-the-art methods.
• Achieved a 95th percentile Hausdorff distance of 5.82 pixels.
• Attained a registration accuracy of 81.2%.

Conclusions:

• CrossModalNet effectively addresses the challenge of low-quality intraoperative DR images in DR-DRR registration.
• The proposed method significantly improves registration accuracy in image-guided radiotherapy.
• The developed technique offers a promising solution for enhancing IGRT precision.