| Image segmentation algorithms identify and delineate objects of interest in an image. Currently, the process to ensure the segmentation is correct is slow and time consuming, as evaluation is done manually. This thesis proposes a testing workflow that can be applied for testing image segmentation software. Pairs of segmentations from different versions of the software are compared using measures defined in literature and this data is used to train classifiers to identify if the pair is consistent. If the classifier produced is acceptable, the evaluation process can be automated. Otherwise, another version of the software must be produced, evaluated manually and added to the training data in order to achieve an acceptable classifier. The case study demonstrates the application of the workflow using software that segments tumors from 3D MRI scans. The workflow is then extended further to show a cost effective way of accelerating the learning process by producing artificially created segmentations in order to create a large training set for machine learning without requiring multiple versions of the software. This thesis also presents a framework for automating the testing workflow process. |
