Cardiovascular disease (CVD) is a major cause of death in the Western world. As such, timely and reliable diagnosis of CVD is a primary requirement in the clinical setting. Manual analysis of multi-dimensional cardiac images (usually regarding the left ventricle) is time-consuming and prone to inter/intra observer variability. Automatic segmentation algorithms are a promising solution to alleviate this issue. Within the model-based segmentation domain, a popular strategy considers the entire segmentation target as a single entity. Although intuitive, this modus operandi suffers from significant practical limitations. One notable example is the requirement of significant training data, due to the difficulty of modelling 3D or 3D+time structures, that exhibit complex spatial and temporal deformation. This thesis investigates an alternate modelling strategy that is adaptable to changes in structure, and scalable with respect to dimensionality. The major contributions presented in this thesis result from investigation into 3D+time cardiac left ventricle segmentation using the proposed approach. The first contribution explores whether all components of a segmentation target need to be explicitly and simultaneously modelled (contour coupling). The second investigates whether complex biological structures can be dimensionally subdivided for modelling and later unified for segmentation (spatio-temporal modelling). The final major contribution determines whether all training data, specifically in a multi-dimensional scenario, is categorically required to construct practical models for accurate segmentation (segmentation framework). Comprehensive evaluation of the proposed method demonstrates that modelling only the crucial components of the segmentation target, while enforcing non-rigid a priori constraints at segmentation-time, allows the proposed method to adapt to configurations outside the training set. It is also illustrated that modelling dimensional variation separately alleviates excessive training requirements and aligning difficulties when compared to the standard unified-modelling approach. In conclusion, this thesis presents a compelling argument for critically evaluating the physical and dimensional structure of the segmentation target to determine the bestsuited modelling strategy. With respect to 3D+time cardiac left ventricle segmentation, the logic of sub-dividing the modelling task into component parts is soundly supported by theoretical and experimental evidence. Finally, a comprehensive segmentation framework is presented to accurately model, and segment, the complex spatial and temporal dynamics of the cardiac structure