QUALITY ASSURANCE USING OUTLIER DETECTIONS FOR CEREBELLAR LOBULE SEGMENTATION
MetadataShow full item record
The cerebellum plays an important role in motor control and cognitive activities. In recent years, several methods have been proposed for automatic cerebellum parcellation. Usually, the segmentation accuracy is evaluated by comparing with manual delineations directly. However, such comparison is impractical in real segmentation scenarios where no manual delineations are available. What is worse, when a segmentation software fails to give an accurate result, the failed segmentation will inevitably bias further studies. Therefore, there is need for an automatic approach that can detect segmentation failures and guarantee the quality of segmentation results. The thesis has two main focuses: evaluating and validating a new approach for cerebellar lobule segmentation and designing an automatic approach for the Quality Assurance (QA) of a cerebellar lobule segmentation pipeline. In the first part of the thesis, we formulate the task of QA and introduce several metrics for evaluating the performance of segmentation software in medical image analysis. We then evaluate a newly proposed cerebellar lobule segmentation software using the introduced metrics. Statistical results show that the segmentation software can give reliable cerebellar lobule segmentation results in a reasonable amount of time while sometimes the software has segmentation failures. The second part of the thesis focuses on automatic QA using outlier detection methods. We introduce a new approach that can automatically detect segmentation failures in a set of segmentation results. The proposed QA approach analyzes all the important processing steps of a segmentation software. In addition, the proposed method provides a general framework of QA that can be modified and applied to other image processing software. Experiments were done on two datasets including healthy controls and subjects with disease. Quantitative results show that the proposed QA method achieves both high sensitivity and high specificity in outlier detection. Qualitative results show that the method can find abnormalities in a set of segmentation results, which should give researchers clues about how their segmentation algorithms perform on a new dataset without ground truth.