Statistical shape models are one of the most important technologies in medical image analysis. With this technology, the computer learns the characteristic shape variations of an organ. The model resulting from this analysis may then be used in implant design, image analysis, surgery planning and many other fields. In this course, we will give an introduction to statistical shape models, from the unifying perspective of Gaussian processes. Besides presenting an overview of the main theoretical ideas underlying these models, we will show how these methods can be applied in practice. In tutorial sessions, the course participants will build their own shape model and apply it for the reconstruction of partial surfaces, using the open source software Scalismo.

In this talk I will present computational and analytical tools we have been developing at University of Utah for the analysis of anatomical image ensembles that are designed to capture changes in anatomy. The fundamental analytical framework we have been using is that of regression analysis where the dependent variable is the anatomical configuration while the independent variable is application domain specific. I will exemplify the application of this general methodology to various medical imaging applications ranging from the analysis of Internal Organ Motion as imaged in 4D respiratory correlated CT imaging spanning few minutes to that of the study changes in brain anatomy associated with normal aging and neurodegenerative diseases such as Alzheimer’s spanning decades.

From price to reproducibility to ease of collaboration and dissemination, open source software is becoming increasingly important for scientific research. This is especially true in the biomedical domain. Having the newest and most advanced methodologies freely available can accelerate research efforts and positively impact millions of lives. In this talk I will take a single data set and demonstrate how to perform various shape analysis tasks using SlicerSALT, an open-source package for dissemination of advanced shape analysis techniques.

Artificial intelligence (AI) has recently been successful applied to many applications, including medical imaging. The basic steps of medical imaging AI development are: 1. Find a clinical unmet need. 2. Define data cohort. 3. Collect medical images and clinical info. 4. Data cleaning and identification. 5. Data set annotation. 6. AI model building, validation and testing. Our team is composed of well-experienced cardiologists and radiologists from various medical centers in Taiwan, as well as researchers specialized at AI/ database management/imaging processing/law of science and technology from different universities. The goal is to build a multi-modality coronary artery disease-based imaging bank, including coronary computed tomographic angiography (CCTA), single photon emission computed tomography (SPECT), invasive coronary angiography (ICA), intravascular ultrasound (IVUS), optical coherence tomography (OCT) and fractional flow reserve (FFR). In this presentation, I will talk about our experience in building a large imaging database for AI research and our preliminary results.

We will follow the story of a patient through one of the most significant milestones of their life: a neurosurgical procedure. Translational digital technologies are used throughout the patient pipeline, from engagement and education to diagnosis, surgical decision support, intraoperative performance augmentation, patient monitoring, recovery, and rehabilitation. Didactic lectures introducing the state-of-the-art in each topic area will be accompanied with interactive engineering exercises supported by provided Jupyter notebooks. At the end of the course, students will have an appreciation of many new ways their work in machine learning can impact clinical practice at the level of an individual patient.

Course Pre-request: Please familiarize yourself with iPython/Jupyter notebooks and come familiar with their use. They will be made available on github prior to the course.

Manifold parameterizations have been applied to various fields of commercial industries. Several efficient algorithms for the computation of triangular surface mesh parameterizations have been proposed in recent years. However, the computation of tetrahedral volumetric mesh parameterizations is more challenging due to the fact that the number of mesh points would become enormously large when the higher-resolution mesh is considered and the bijectivity of parameterizations is more difficult to guarantee. In this talk, we develop a novel volumetric stretch energy minimization algorithm for volume-preserving parameterizations of simply connected 3-manifolds with a single boundary under the restriction that the boundary is a spherical area-preserving mapping. In addition, our algorithm can also be applied to compute spherical angle- and area-preserving parameterizations of genus-zero closed surfaces, respectively. Several numerical experiments indicate that the developed algorithms are more efficient and reliable compared to other existing algorithms. Numerical results on applications of the manifold partition and the mesh processing for three-dimensional printing are demonstrated thereafter to show the robustness of the proposed algorithm.

Recent advances in computer vision, including leaps in machine learning systems, fuel cutting edge research on contextual and task-aware computer assistance solutions that streamline workflows while catering to the physicians' needs to enable improved clinical decision making. In this talk, I will highlight some of our recent work aiming to push the boundaries in navigated sinus endoscopy. I will focus on self-supervised learning from monocular video without photometric constancy, demonstrating applications in monocular depth estimation, correspondence finding, and photo-realistic dense reconstruction.

Recent medical image registration methods based on deep neural networks have seemly converged to so-called end-to-end learning approaches, in which the networks are trained to directly predict spatial transformation between a given pair of unprocessed images. Besides fast inference that enables sub-second volumetric registration, some have shown superior accuracy over the classical methods. I will first give a technical overview of the classical medical image registration and the state-of-the-art deep-learning-based examples, including the unsupervised and the (weakly-) supervised methods. I will conclude this talk with an introduction to our latest development using a more data-driven attempt to replace the transformation-predicting alternatives.

Taiwan Society for Industrial and Applied Mathematics