Achieving accurate colour image segmentation in 2D and 3D with LVQ classifiers and partial adaptable class-specific representation

Adaptable Class-Specific Representation (ACSR) has previously been used as a solution to the problem of segmentation near edges in 2D colour images. Sampling windows of fixed shape used in many segmentation approaches cause misrepresentation of texture classes. ACSR greatly reduces this problem, based on simple templates, resulting in accurate semi-automatic segmentation. The price of accuracy in ACSR is high processing overhead. We introduce an initial segmentation step using a faster fixed-shape window sampling and Learning Vector Quantization, and apply ACSR only at edge point. Processing speed is significantly increased without compromising segmentation accuracy. ACSR segmentation is particularly interesting for medical applications where correct shape and size is important. We extend the ACSR framework to true 3D volume segmentation. 3D information is used for classification at all sampling points, producing better results than per slice pseudo-3D segmentation. Colour volumes based on the Visible Human Project are used to demonstrate the approach. We conclude that ACSR can produce accurate segmentation in colour 2D images and 3D volumes, and that partial ACSR can significantly reduce processing overhead without losing segmentation quality.