Pathology Detection on Medical Images Based on Oriented Active Appearance Models

In this paper, we propose a novel, general paradigm based on creating a statistical geographic model of shape and appearance of normal body regions. Any deviations from the normality information captured in a given patient image are highlighted and expressed as a fuzzy pathology image. We study the feasibility of this idea in 2D images via Oriented Active Appearance Models (OAAM). The OAAM synergistically combines AAM and live-wire concepts. The approach consists of three main stages: model building, segmentation, and pathology detection. The model is built utilizing image data from normal subjects. The model currently includes shape and texture information. A variety of other information (functional, morphometric) can be added in the future. For segmentation, a novel automatic object recognition method is proposed which strategically combines the AAM with the live-wire method. A two level dynamic programming method is used to do the finer delineation. During the process of segmentation, a multi-object strategy is used for improving recognition and delineation accuracy. For pathology detection, the model is first fit to the given image as best as possible via recognition and delineation of the objects included in the model. Subsequently, a fuzzy pathology image is generated that expresses deviations in appearance of the given image form the texture information contained in the model. The proposed method was tested on two clinical CT medical image datasets each consisting of 40 images. Our preliminary results indicate high segmentation accuracy (TPVF>97%, FPVF<0.5%) for delineating objects by the multi-object strategy with good pathology detection results suggesting the feasibility of the proposed system.