Transfer Learning for Nonrigid 2D/3D Cardiovascular Images Registration.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Surgery

Background:

• Cardiovascular image registration combines preoperative 3D computed tomography angiography (CTA) and intraoperative 2D X-ray/digital subtraction angiography (DSA) for minimally invasive vascular interventional surgery (MIVI).
• Convolutional neural network (CNN) regression models offer fast and accurate registration but struggle with inter-patient variability.
• Patient-specific CNN training requires extensive datasets, limiting generalizability and clinical application.

Purpose of the Study:

• To evaluate the efficacy of transfer learning (TL) for improving 2D/3D deformable cardiovascular image registration.
• To adapt CNN models trained on one patient's data for accurate registration in other patients.
• To reduce the data requirements for accurate image registration using TL.

Main Methods:

• Utilized transfer learning by optimizing frozen weights in convolutional layers to identify common feature extractors.
• Applied TL to a nonrigid registration model for 2D/3D cardiovascular images.
• Compared the performance of the TL-enhanced model against a non-TL model and traditional intensity-based methods.

Main Results:

• The transfer learning approach significantly reduced the required training dataset size to 200 images for accurate registration.
• The nonrigid registration model incorporating TL demonstrated superior performance compared to models without TL.
• The TL-enhanced model outperformed traditional intensity-based registration methods for deformable cardiovascular images.

Conclusions:

• Transfer learning is a viable and effective strategy for enhancing the accuracy and efficiency of 2D/3D cardiovascular image registration.
• TL overcomes the limitations of patient-specific CNN training, enabling broader application in MIVI.
• The proposed TL-based nonrigid registration model offers a promising solution for complex cardiovascular interventions.