Automated Lung Segmentation and Image Quality Assessment for Clinical 3D/4D Computed Tomography

Four-dimensional computed tomography (4DCT) provides not only a new dimension of patient-specific information for radiation therapy planning and treatment but also a challenging scale of data volume to process and analyze. Manual analysis using existing 3D tools is unable to keep up with vastly increased 4D data volume, automated processing and analysis are thus needed to process 4DCT data effectively and efficiently. In this work, we applied ideas and algorithms from image/signal processing, computer vision and machine learning to 4DCT lung data so that lungs can be reliably segmented in a fully-automated manner, lung features can be visualized and measured on-the-fly via user interactions, and data quality classifications can be computed in a robust manner. Comparisons of our results with an established treatment planning system and calculation by experts demonstrated negligible discrepancies (within ±2%) for volume assessment but one to two orders of magnitude performance enhancement. An empirical Fourier-analysis-based quality measure delivered performances closely emulating human experts. Three machine learners are inspected to justify the viability of machine learning techniques used to robustly identify data quality of 4DCT images in the scalable manner. The resultant system provides tools that speeds up 4D tasks in the clinic and facilitates clinical research to improve current clinical practice.