Interactive texture analysis in chest CT scans

Abstract

Annotation of structures or abnormal areas in medical images is necessary for various clinical and research purposes, but it is a labor-intensive task. This thesis describes an interactive system for chest CT scans, developed to ease annotation. First, the structure of interest, either the chest or the lungs, is segmented automatically. This structure is divided into smaller volumes of interest (VOIs) containing one type of texture, which are automatically labeled. An observer interactively refines these labels. This can be done by changing the label of incorrectly labeled VOIs one by one, or by training a classifier. In the latter case, the observer is shown an axial slice of the scan with the automatically assigned VOI labels. The observer corrects the labeling errors and the VOIs reviewed by the observer are used to train a classifier. This classifier classifies the VOIs in a second axial slice and the observer again corrects any mistakes. The VOIs in the second slice are added to the training data and the classifier is retrained. The process of correction, retraining, and classification is ended when a predefined percentage of the structure of interest is annotated, or when the user thinks the classifier is trained well enough. After that, the classifier is trained one last time and all remaining VOIs are classified. The observer inspects all annotations to correct any remaining errors, after which he or she saves the annotation results. Interactive classification was applied to CT scans of the lungs of patients with interstitial lung disease, using software to simulate observer behavior. Normal lung tissue and seven types of abnormal lung tissue were distinguished. On average, the labels of 80% of all VOIs were correctly predicted. Using further optimization strategies, a median accuracy of 88% was reached. In addition, interactive texture analysis was used to segment the lungs in CT scans of pigs, mice, and human patients. On average, interactive lung segmentation took less than 9 minutes of user interaction, for changing the labels of 2.0% of the VOIs in a scan. The resulting segmentations showed a good correspondence to manual lung segmentations. Finally, interactive lung segmentation and texture analysis in chest CT scans were applied to chest CT scans of intensive care patients, who required mechanical ventilation. In current clinical practice, the amount of air delivered by the mechanical ventilator is determined based on the height of the patient and on the presence or absence of lung injury. We compared CT measurements of total lung capacity and normally aerated lung to estimations based on patient height. Total lung capacity and normally aerated lung volume could not be reliably estimated with the current clinical approach. The described interactive approach might be a first step towards personalized prediction of optimal ventilator settings.