Transfer Learning for Rigid 2D/3D Cardiovascular Images Registration

Cardiovascular image registration is an essential approach to combine the advantages of preoperative 3D computed tomography angiograph (CTA) images and intraoperative 2D X-ray/ digital subtraction angiography (DSA) images together in minimally invasive vascular interventional surgery (MIVI). Recent studies have shown that convolutional neural network (CNN) regression model can be used to register these two modality vascular images with fast speed and satisfactory accuracy. Because of the large differences in the vascular architecture of different patients, a CNN regression model trained on one patient often cannot be applied to another. To overcome this challenge, we proposed a transfer learning based CNN regression model which can be transferred from one patient to another with only tiny modifications. The registration error of our proposed method can reach less than 1 mm or 1\(^ \circ \) when a trained model is fine-tuned with only 200 images of the target patient in about 150 s. We tested the transfer ability of our method with images from various patients suffering different cardiovascular disease and confirm the effectiveness of our method. Deformation of cardiac vessels was not considered in this rigid registration model and non-rigid cardiovascular registration model will be developed in our future work to improve the registration accuracy of \(t_z\).