Image segmentation of spectral computed tomography images is important for various pre-clinical studies when measuring and analysing a region of interest. However, accurate segmentation is challenging to achieve with automatic segmentation algorithms. This is because spectral CT is a new imaging modality with a limited number of available datasets or corresponding labels, varying number of energy channels per scan, and vast differences in the shape and size of an object due to the pre-clinical nature of the studies. This thesis demonstrates the need and development of accurate interactive image segmentation algorithms.

An ideal interactive segmentation algorithm should be computationally fast, accurate, and have minimal user interaction. Keeping these characteristics as a context, a region based interactive segmentation algorithm using a bag of features approach is proposed for 2D slices. During evaluation, this approach had shown better performance than some commonly used methods. However, even though bag of features for feature extraction seemed effective, this approach doesn’t exploit the spatial information provided by the user.

Hence, a new spatially aware approach is introduced to explicitly include user cues in the segmentation process. It also investigates smooth segmentation boundaries by following two stage conditional random fields method during the inference stage. Experimental results showed that this approach outperformed traditional interactive segmentation algorithms and also the incorporating spatial cues did indeed improve the segmentation results. This algorithm is also implemented as a software tool to be used by researchers and clinicians.

The proposed spatially aware approach is then modified and extended to work with 3D volumes in a slice-wise manner, to make sure that the user interaction is propagated across all slices. Compared to the direct extension of the algorithm from 2D to 3D, the slice-wise approach resulted in well connected segmentation regions with fewer false positive and false negative regions.

Finally, using image segmentation as a primary step, a software tool is proposed and implemented to assess bone health. It is developed to overcome the challenges posed by ImageJ’s bone analysis plugin. The experiments done using both synthetic and real datasets showed that the proposed tool is proven to be better in terms of speed and accuracy.