Purpose: Ground Glass Opacity (GGO) is a radiologic feature which appears as a pulmonary opacity with vague boundaries without obscuring the underlying vascular and pulmonary markings on a Chest Computed Tomography (CT) scans. GGO is a non-specific sign seen in various pathologies, most commonly: alveolar inflammation, infection, hemorrhage, edema or cancer. Relative to more discreet findings, GGO is often considerably more fine and subtle and thusly at times overlooked.
We present an automated method for detection of GGO in CT scans based on a Fully Convolutional Neural Network.
Material and methods: We utilized segmentation of axial CT reconstructed images to reduce the number of CT studies required as training data to obtain high accuracy of GGO detection (96.9%) using a Deep Learning (DP) technique. We explore two architectures of Fully Convolutional Neural Networks: U-Net and Fully Convolutional DenseNet. DenseNet-like network is first applied to the Medical Imaging domain and achieves superior detection accuracy due to a higher layer connectivity within a network. We report results of GGO binary classification per axial slice and measurement of slice segmentation goodness (Dice score). The algorithm is constructed to be applicable to any Chest CT scan, allowing for variations in data acquisition protocols such as inspiration/expiration imaging and technical acquisition variations which may result in the appearance of the lung tissue.
Results: The best accuracy was achieved on the deepest FC-DenseNet (with 103 convolutional layers) - 96.9% (with specificity 98.3% and sensitivity 80.5%). Higher accuracy is obtained from a deep FCN with skip connections, which follows the principle that convolutional networks with skip connections are much easier to optimize than deep CNNs. The best Dice score, 72.5%, is achieved by U-Net with 97 convolutional layers, which is impressive considering the level of inconsistency in GGO tagging due to high vagueness in the opacity boundaries.
Conclusions: We accomplished an end-to-end approach for Ground Glass Opacity detection using Fully Convolutional Neural Networks. We explored several architectures of varying depth, U-Net- and DenseNet-like FCNs, and, as expected, the deepest networks showed the best performance. The best accuracy we achieved is 96.9%, which is promising, given the complexity of GGO detection and represents a significant improvement on similar prior endeavors.
