[home] [Personal Program] [Help]

---

Session: Poster session 2 (Odd numbers)
Session starts: Friday 27 January, 10:00
Presentation starts: 10:00

---

Dieuwertje Alblas (University of Twente)
Christoph Brune (University of Twente)
Kak Khee Yeung ()
Jelmer Wolterink ()

Abstract:
Personalised 3D vascular models are valuable for diagnosis and treatment planning in patients with cardiovascular diseases. Acquiring explicit representations of these models from image data is time-consuming, and often requires pre-processing for downstream tasks. We propose to represent these surfaces implicitly by the zero levelset of their signed distance function in an implicit neural representation. For training, this network only requires a set of points on the vessel’s surface. The network represents the surface at any resolution with high accuracy. We briefly describe the method and its benefits. The signed distance function (SDF) of a surface M is defined as: SDF_M (x)= {█(-d(x, M)@0@d(x, M) )┤ ■(x inside M @x on M @x outside M) Where d(x, M) is the distance to the surface. The SDF is a continuous function that can be evaluated at any x in R^3. We use a neural network f(x;θ) to approximate the SDF [1, 2]. Similar to the SDF, the network takes in coordinates x and outputs the local SDF value. For training, we used a set of points X⊂R^3 on the vessel’s surface. We assessed the robustness and accuracy of this method by reconstructing 3D mesh models of the abdominal aorta. We can reconstruct the vessels with a Dice Similarity Coefficient of 0.94 with as few as 200 points. Moreover, we reconstructed the SDFs of layered structures in the vessel wall with a single neural network. The neural network captures the nested topology of the shapes, resulting in zero unwanted extrusions of inner structures, in contrast to representing each shape with a separate neural network. Lastly, smoothly connecting multiple vessels into a vascular tree is cumbersome for mesh models. In our approach, multiple vessels can be easily combined into a smooth vascular tree using a smooth minimum operation on their SDFs. We demonstrated a method that reconstructs smooth, watertight vessels, based on a set of points. This can be readily adapted into an annotation pipeline to acquire personalized 3D vascular models from image data. Smooth blending of these vessels makes the obtained vascular models directly suitable for downstream tasks, e.g. computational fluid dynamics.