4D Multi object segmentation based on MR image sequences - Medical application for evaluation of myocardial differences in shape and function after infarction

Despite years of medical research, myocardial infarction remains one of the leading causes of premature mortality in Western industrialized nations. Patients' quality of life and prognosis depend significantly on whether cardiac muscle function recovers after an acute myocardial infarction or whether a contraction disorder persists and leads to progressive deterioration of cardiac function with cardiac remodeling. Early identification of patients at risk for remodeling is important for promptly initiating effective therapies to prevent remodeling.
The aim of this proposal is to develop methods for the automatic detection, quantitative description, and prediction of myocardial remodeling in clinical spatiotemporal MRI datasets. For this purpose, a workflow will be developed that includes automated preprocessing, model-based 4D segmentation and motion estimation in the image data, as well as the extraction of clinically relevant parameters, their analysis, and suitable visualization. These methods will be developed and evaluated using a comprehensive set of baseline and follow-up MRI datasets from infarction patients acquired using clinically standard acquisition parameters, as well as image data from healthy subjects. A total of over 360 anonymized MRI datasets with various MRI sequences (including cine MRI, LGE MRI, T2w MRI) as well as manual contouring of the cardiac structures are available for the project.
A central aspect is the development of a model-based approach for integrated segmentation and motion estimation of the left and right ventricles, which incorporates prior knowledge of both the shape and form variability as well as the movement of the heart. Another central topic is the quantitative analysis and classification of clinical image data based on a patient cohort. Learning-based algorithms will be used to extract relevant parameters from the image data and the derived shape and motion information and – in conjunction with the model-based segmentation and registration procedure – enable automated detection and prediction of myocardial remodeling.

Heart attacks and their degenerative consequences are one of the main causes of premature mortality in Western countries. To adequately investigate the effects of myocardial infarction on the geometry and motion of the heart, a variety of image-based analysis methods are used in clinical settings. These methods generally require accurate segmentation and motion estimation of individual cardiac regions (ventricles, atria, etc.). This project investigated cardiac segmentation and motion estimation in clinical cine MRI image data, which spatially and temporally depict the heart during a cardiac cycle.
For this purpose, a model-based 4D method for integrated segmentation and registration was developed. Based on the model information, this method is capable of generating spatially and temporally consistent segmentations of the various cardiac regions and also provides surface-based motion information. In practice, however, the accuracy and generalization capability of such a model-based segmentation and registration method depends crucially on the quality and quantity of the training data available for training the model—i.e., manually segmented image data. Since the availability of such training data is generally limited in a medical context, two methods were developed during the project that enable the generation of flexible shape, appearance, and deformation models with high generalization capability based on a small amount of training data.
The basic idea of ​​the developed approaches is the application of the locality principle, known from physics, to the objects to be modeled. This means that distant object structures are assumed to have only an indirect influence on the local shape or appearance. The developed approaches are naturally suited for multi-object segmentation and can be embedded in a multi-level approach, which enables the integrated representation of global and successively more local shape and appearance variations. Various experiments have shown that the developed approaches are capable of generating better generalizing models for small training populations than state-of-the-art methods, which in practice also translates into improved segmentation results. The high generalization capability further enables the application of this approach for model-based data augmentation for CNN-based learning methods. Thus, using a small amount of training data, a shape and appearance model was generated that generates any number of simulated data sets, which then allow the training of CNNs for medical image registration. The trained CNNs can then be used, for example, to quickly and accurately estimate heart motion in spatiotemporal image data.