Signed Distance Field based Segmentation and Statistical Shape Modelling of the Left Atrial Appendage

TL;DR

This work addresses the need for quantitative, robust LAA morphology descriptors to relate shape to stroke risk in atrial fibrillation. It introduces a fully automatic pipeline that regresses a Signed Distance Field ($SDF$) from CT images to create smooth LAA meshes, followed by non-rigid registration to a template and construction of a Statistical Shape Model ($SSM$) via PCA, with shape clustering via Gaussian Mixture Models. The approach achieves high segmentation accuracy (Dice scores in the high 90s) and sub-millimeter mesh fidelity, while revealing that the LAA morphology can be described by a small number of principal modes and two distinct shape clusters (chicken-wing vs non-chicken-wing). The method enables scalable, user-independent morphological descriptors suitable for large-scale studies linking LAA shape to ischemic stroke risk and to in-silico haemodynamics analyses.

Abstract

Patients with atrial fibrillation have a 5-7 fold increased risk of having an ischemic stroke. In these cases, the most common site of thrombus localization is inside the left atrial appendage (LAA) and studies have shown a correlation between the LAA shape and the risk of ischemic stroke. These studies make use of manual measurement and qualitative assessment of shape and are therefore prone to large inter-observer discrepancies, which may explain the contradictions between the conclusions in different studies. We argue that quantitative shape descriptors are necessary to robustly characterize LAA morphology and relate to other functional parameters and stroke risk. Deep Learning methods are becoming standardly available for segmenting cardiovascular structures from high resolution images such as computed tomography (CT), but only few have been tested for LAA segmentation. Furthermore, the majority of segmentation algorithms produces non-smooth 3D models that are not ideal for further processing, such as statistical shape analysis or computational fluid modelling. In this paper we present a fully automatic pipeline for image segmentation, mesh model creation and statistical shape modelling of the LAA. The LAA anatomy is implicitly represented as a signed distance field (SDF), which is directly regressed from the CT image using Deep Learning. The SDF is further used for registering the LAA shapes to a common template and build a statistical shape model (SSM). Based on 106 automatically segmented LAAs, the built SSM reveals that the LAA shape can be quantified using approximately 5 PCA modes and allows the identification of two distinct shape clusters corresponding to the so-called chicken-wing and non-chicken-wing morphologies.

Figures (8)

• Figure 1: A) Schematic illustration of the human heart together with four examples of Left Atrial Appendage (LAA) morphologies from different subjects in our data set. RA: right atrium, RV: right ventricle, LA: left atrium, LV: left ventricle. B) Examples of commonly used measures of LAA morphology
• Figure 2: Overview of the full processing pipeline. A region-of-interest is detected around the left atrium (LA) and a signed distance field (SDF) is regressed based on it. The zero-level isosurface is extracted and the left atrial appendage (LAA) is decoupled from the LA. All LAA shapes are registered to a common template (Source) using volumetric registration and a statistical shape model (SSM) is built. The SSM enables us to investigate the main modes of variation, generate realistic synthetic shapes and conduct unsupervised clustering.
• Figure 3: Point correspondence is established between the source and each target using volumetric registration (Vol.Reg). The two surfaces are pre-aligned using an iterative closest point (ICP) algorithm, before the SDFs are registered and the source is transformed with a B-spline transformation. All points on the transformed source are projected to the target mesh using markov random field regularization of the correspondence field (MRF cor.). Based on the neck points defined on the original source, we find the best fitting plane that decouples the LAA from the LA. The method for estimating point correspondence is repeated, this time using the decoupled LAAs for Source and Target.
• Figure 4: Surface visualization of results from our voxel-wise-regression (VWR), manual expert annotation and voxel-wise-classification (VWC) without signed distance field regularization. The chosen examples correspond row-wise to the best, median and worst results evaluated based on dice-score. The baseline VWC with SDF regularization is visually equivalent to VWC without regularization and therefore omitted.
• Figure 5: Manual expert annotation (mesh) and our proposed method (full surface) from all 20 LAAs from the testset. The results are ranked based on Dice-score with the best result in the top left and the worst in bottom right.