This talk gives an overview of the different components required for a fully automated 3D segmentation pipeline based on statistical shape models (SSMs) and machine learning. Starting from the construction of mesh-based statistical shape models, over initialization and appearance modeling, up to the local adaptation strategy, different approaches are presented and discussed. The focus is on practical applicability of all techniques and should enable newcomers to get a good starting point for a first SSM pipeline, as well as give more experienced participants ideas for further optimizations and refinements.

Rapid advances in image acquisition and shape capturing technology provide continuous or time-discrete 3D volumetric images and/or surfaces, motivated by the notion that dynamic spatiotemporal changes may provide information not available from snapshots in time. Capturing properties of time-varying 3D objects requires novel methods to make use of the inherent correlation and causality of repeated measurements. This talk will discuss development of advanced 4D analysis methodologies that carry the notion of linear and nonlinear regression, now applied to complex, high-dimensional data such as image-derived shapes and structures. We will address concepts for regression of 4D shapes from time-discrete data, enforcing temporal smoothness, and will discuss work in progress towards statistical analysis of 4D shape trajectories. Comparison of shape change trajectories will also include the notion of time warp for accommodating variations in time-change functions. The proposed methodology is based on the framework of “currents” and thus overcomes limitations of common point-to-point correspondence. Our research is driven by challenging medical image analysis problems, where clinical assessment routinely uses terms such as development, growth trajectory, aging, degeneration, disease progress, recovery or prediction. This terminology inherently carries the aspect of dynamic processes, suggesting that snapshots in time and cross-sectional analysis in populations are not sufficient. We will show examples from ongoing clinical studies such as analysis of early brain growth in healthy and at-risk subjects, of modeling neurodegeneration, and of modelling bone shape developments after surgery.

In human anatomy, multiple structures are interrelated and hierarchically organized. We present computational anatomy modelling approaches which incorporate these interrelations and hierarchization to perform automated segmentation and identification of patient-specific anatomy of abdominal organs and musculoskeletal structures from 3D data. Further, the approaches are extended so as to address automated surgical planning and diagnostic assistance.

Structural shape analysis has become an important tool for local quantification of shape differences between groups of subjects, shape changes across time or for the discrimination of patient groups. This presentation will focus on methods available within the NA-MIC shape toolkit that encompasses object-specific spherical harmonic descriptions, derived medial descriptions, and group-based particle descriptions. Extensions for longitudinal data and complex surface data are presented as well.

Statismo is an open source framework for statistical shape modeling. It provides functionality for all shape modeling tasks, from model-building to shape analysis. This tutorial will introduce the basic principles behind statismo and shows on concrete examples how it can be used in practical applications. In a first part, we will introduce the basic philosophy and design principles of statismo. The second part shows with concrete code examples of how statismo can be used for building shape models and model fitting. In the last part we show how the concept of Gaussian process models in statismo can be used to formulate shape priors that are far more powerful than standard PCA models.