Ground-glass opacities (GGO) and consolidation are findings typically observed in pneumonia and other acute lung diseases. For radiologists, segmentation and analysis of GGO and consolidation on chest computed tomography (CT) images are critical to quantify the disease severity and assess stages of recovery. However, these disease areas are difficult to detect, differentiate and segment on CT scans as these usually do not have clear boundaries. Furthermore, radiologists’ manual segmentations of lung lesions are a highly subjective and time-consuming task considering that a high-resolution Chest-CT study comprises hundreds of images. In this paper, we present a deep learning approach that automatically segments GGO and consolidation in Chest-CT studies. Our approach first segments the region of interest (ROI), lung parenchyma, achieving a Dice similarity coefficient of 0, 97(±0.02) and a Jaccard coefficient of 0, 95(±0.04). Subsequently, to train our deep learning model with a small dataset of GGO and consolidation CT images, the model training is performed using an active learning approach. The results for semantic segmentation achieve a DSC of 0, 81(±0.12) and a Jaccard coefficient of 0, 70(±0.15) outperforming the supervised learning approach. Our results suggest that active learning reduced the amount of labeled data required for medical image segmentation without significant loss of accuracy.

Keywords: Consolidation, ground-glass opacity, lung infections, lung segmentation, deep learning, active learning

For more information read the full article:

A deep learning approach for automatic segmentation of ground-glass opacities and consolidation in small Chest-CT dataset