Ontologies are an effective means to formally specify and constrain knowledge. They have proved their utility in various data mining applications, especially in annotating text to render it machine interpretable. More challenging research perspectives arise when ontologies are used to annotate images where the information is encoded in numeric pixel values rather than in natural language. Current approaches to bridge the gap between the pixel-based foundational representation and high level image semantics include the utilization of taxonomies describing 2D spatial relations between the depicted objects and hence linking image features with semantics. To this end we present a novel ontological approach that formalizes concepts and relations regarding image representations for medical image mining. It provides descriptors for pixels, image regions, image features, and clusters. It extends previous approaches by including assertions of spatial relations between clusters in multidimensional feature spaces. The relational assertions enable the linkage between a given image, image region and feature(s) to the object they represent. The proposed approach is more general than most current approaches and can be easily extended to support multimodal data mining.