“Blob” analysis of biomedical image sequences: A model-based and an inductive approach

Abstract

New directions in robotics and computer vision indicate that useful behaviours of artificial systems can be achieved with simple reactive control strategies and with a task dependent representation of incoming data.This philosophy has been used in the design and implementation of an off-line system for automatic object detection and delineation in biomedical image sequences, where the task is to estimate object area vs. time. The image is represented with feature vectors on a coarse resolution and scale, and the image is processed using a scene model and a procedural model. The scene model expresses relations between objects in the scene, the objects being represented by regions, or “blobs”. The objects are labeled with a relaxation labeling algorithm, and a constraint satisfaction algorithm. The procedural model — a finite state automaton — expresses the different processing paths due to possible model-data mismatches, including processing on a higher resolution and the use of alternative scene models. The automaton comprises a top level supervisor as well as a lower level reactive control mechanism. The control mechanism and the scene model are exchangeable between different applications. Images in the sequence are re-processed if the objects of interest in consecutive images do not overlap.As a comparison, unsupervised learning with Kohonen's self-organizing feature maps were used to train the system to perform segmentation and delineation. The feature map was trained with a low dimensional feature vector randomly sampled from a population of representative images. During the association phase, for each feature vector, the map nodes are searched for the best matching node, and the nodes that correspond to the desired object are grouped into larger regions. The object is delineated from the largest region retrieved from the map.Two domains have been analyzed by the system, ultrasound image sequences of the heart and gamma camera sequences of the heart.KeywordsFeature VectorGamma CameraFinite State AutomatonImage UnderstandingScene ModelThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.