Recent progress in fully-automated image segmentation has enabled efficient extraction of clinical parameters in large-scale clinical imaging studies, reducing laborious manual processing. However, the current state-of-the-art automatic image segmentation may still fail, especially when it comes to atypical cases. Visual inspection of segmentation quality is often required, thus diminishing the improvements in efficiency. This drives an increasing need to enhance the overall data processing pipeline with robust automatic quality scoring, especially for clinical applications. We present a novel quality control-driven (QCD) framework to provide reliable segmentation using a set of different neural networks. In contrast to the prior segmentation and quality scoring methods, the proposed framework automatically selects the optimal segmentation on-the-fly from the multiple candidate segmentations available, directly utilizing the inherent Dice similarity coefficient (DSC) predictions. We trained and evaluated the framework on a large-scale cardiovascular magnetic resonance aortic cine image sequences from the UK Biobank Study. The framework achieved segmentation accuracy of mean DSC at 0.966, mean prediction error of DSC within 0.015, and mean error in estimating lumen area ≤17.6 mm2 for both ascending aorta and proximal descending aorta. This novel QCD framework successfully integrates the automatic image segmentation along with detection of critical errors on a per-case basis, paving the way towards reliable fully-automatic extraction of clinical parameters for large-scale imaging studies.
750 - 758