A model-guided feedback control system for image segmentation

This dissertation work presents a feedback control mechanism for image segmentation systems. Current model-guided segmentation systems typically use sequential multi-stage search of target objects. They first locate obvious structures and using those key landmarks, subsequently segment finer structures. Therefore, any incorrect or inaccurate intermediate result at one stage may cause subsequent stages to fail.; Applying the concept of feedback control system in control engineering, the proposed mechanism compares the model of desired output to the actual output to detect and then correct errors at intermediate stages of multi-stage image segmentation systems. By modeling the expected output of each stage, different types of errors can be detected and correspondingly different error correction strategies employed.; In this research, a feedback control system was designed, implemented and tested on four different sets of medical imaging problems to show generalizability: hand contour verification in pediatric hand radiographs, distal phalange bone contour verification in hand radiographs, lesion detection in lung CT images, and finally, the separation verification of left and right lungs in CT lung images.; The performance of the feedback control system applied to the hand contour problem was evaluated using clinical images obtained from UCLA clinical PACS. Graders rated the correctness of each image's segmentation result before and after the feedback system was applied. The resulting statistical analysis comparing the performance without and with the feedback mechanism demonstrated that feedback significantly improved the overall correctness of the hand contours. From a total of 76 images, 42 (55%) underwent error detection and correction. Three graders gave approximately 75% of these cases better scores. A similar evaluation methodology was applied to measure the performance of the feedback system for lung lesion detection. The system performed well in identifying lesions in the outer lung boundary with a correctness score average of 4.2 and median of 5 on a scale of 5.; In summary, this research has successfully applied feedback mechanism to the above several multi-stage segmentation systems. This approach has significantly improved the segmentation performance in each case. This demonstrates the usefulness and generalizability of this approach and shows the promise for further development.