A Patient-Specific Registration of Coronary Angiogram-Fluoroscopy by Similarity- Based Transfer Learning

Percutaneous coronary intervention (PCI) is an effective treatment for normalizing blood flow in coronary arteries narrowed by stent implantation. Guidewire insertion during PCI requires considerable precision and expertise, and two X-ray videos, a cine loop of angiography and a real-time video of live fluoroscopy, are used to aid guidewire navigation to the target location without entering the wrong blood vessel branches. However, simultaneously watching and interpreting two radiographic videos warrants increased mental effort and intervention time. Additionally, more contrast agents may have to be injected to verify the insertion. Although deep-learning-based dynamic coronary roadmapping (DCR) has been suggested to provide registered images from two different X-ray sources, existing methods may not be suitable for irregular heartbeats or may vary in effectiveness depending on the patient, posing challenges in time-critical situations. To address these challenges, we propose a patient-specific approach for DCR using similarity-based patient data matching in transfer learning with a residual U-Net. The proposed method leverages the anatomical similarities between newly acquired and pre-acquired angiograms by utilizing principal component analysis and cosine similarity to facilitate efficient transfer learning. Moreover, a residual U-Net architecture that incorporates residual blocks and leaky ReLU activation functions to accelerate patient-specific transfer learning is proposed. These advanced techniques resulted in significantly fast transfer learning of less than five minutes from a pre-trained model, as well as high registration image quality with over 30 dB in peak signal-to-noise ratio, while maintaining a registration error of $1.04\pm 0.19$ mm.