Uncertainty modelling for end-to-end 3D reconstruction of
\ncoronary arteries from 2D X-ray angiography

Coronary artery diseases are one of the main causes of death in Canada. Amongst those, stenoses of the coronary arteries are one of the most predominant. During a percutaneous intervention, a catheter is inserted through the femoral artery and guided towards to heart. During stenting procedures, a stent is guided alongside the catheter. Upon reaching the narrowed artery, the catheter is expanded to give the affected artery a more tubular shape. \n \nX-ray angiography are currently the gold-standard imaging procedure for the guidance of catheters. These images are obtained using X-rays while injecting contrast agents in the patient’s arteries. The main difficulty linked to the use of angiography is linked to the noise contained in the images and the ambiguities created by the projection of the 3D structures onto 2D images. As such, the outcome of the stenting procedures is intimately linked to the experience of the cardiologists. Using a 3D model of the arteries during percutaneous interventions could help alleviate the difficulty encountered during such interventions. \n \nHence, our main objective is to model uncertainty in the context of 3D reconstruction of coronary arteries during percutaneous interventions. Our contributions are three-fold: 1. Coronary artery segmentation on X-ray angiography using uncertainty metrics; 2. Fully customizable coronary artery angiography synthesis; 3. Monocular 3D reconstruction of coronary arteries using mesh deformation networks. \n \nThe first contribution is a novel method to segment coronary arteries in X-ray angiography. This is done using Bayesian Convolutional Neural Networks to also provide a pixel-wise measure of uncertainty regarding the yielded segmentation. This measure of uncertainty is then used alongside a fully-connected neural network designed to provide a threshold for the uncertainty values. The values above the threshold and then deemed too risky to use directly and flagged for the operator while the values below can safely be used for interventions. \n \nOur second objective considers a new method to synthesize coronary artery X-ray angiography. We used a realistic cardio-respiratory simulator to generate fully customizable sequences of coronary arteries. We proposed a new loss function designed to work with CycleGAN to transfer the style of X-ray coronary arteries onto the simulated images. The new loss function is based on a vesselness measure that checks that the topology of the coronary arteries from the input image is respected in the stylized images. Our method allows for the generation of new images for learning purposes or data augmentation purposes by allowing the generation of out-of-distribution data. \n \nFor our last objective, we proposed the single-view 3D reconstruction of coronary arteries from a segmented X-ray angiography. To do so, we used two learning models. One model is trained to extract visual and geometrical features from the segmented image. The other model uses the extracted features to adapt a mesh and gradually make it adapt to the shape of the object to reconstruct. Using a single image for the 3D reconstruction alleviates the need for temporal registration and allows the application of the method to multiple catheterization laboratory configurations. The obtained reconstruction can be used as an additional reference for the guidance of catheters during percutaneous interventions.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.