Zielsetzung
Chronic Obstructive Pulmonary Disease (COPD) is a leading cause of death worldwide, yet early detection and treatment can prevent the progression of the disease. X-ray Computed Tomography (CT) scans of the chest provide a quantitative measure of morphological changes in the lung. The potential of incorporating optimal Window Setting (WS) selection, typically carried out during examination of CT scans for COPD, is generally overlooked in deep learning models. We aim to optimize the detection of COPD with densely connected convolutional neural networks (DenseNets) through implementation of Window-Setting Optimization (WSO) steps.
Material und Methoden
We adapted this routine clinical-workflow procedure for a total of 78 chest CT scans. Training (n=3,392), validation (n=1,114), and test (n=2,688) sets were separated on slice-level basis, and balanced in number of slices representing healthy and COPD patients. Slices were preprocessed to the lung, emphysema, and full-range of Hounsfield unit scale WS. First, the effect of manually-adjusted WS for a plain DenseNet model was examined. Next, with introduction of a WSO layer to the DenseNet model, automated WSO was explored.
Ergebnisse
Repeated inference over 3 runs for the test set showed that the plain DenseNet model trained with images preprocessed to the emphysema WS predicted COPD with a mean slice-level Area Under the Receiver Operating Characteristics curve (AUC) of 0.87+-0.021. Comparatively, t-test analysis demonstrated that preprocessing the images to the lung WS resulted in a significantly worse outcome (AUC=0.74+-0.021, p=0.0029). By automating the WSO process, an optimal WS in the proximity of the emphysema WS was learned, and an AUC of 0.85+-0.041 was achieved. This WSO model showed improvement over the plain model trained with images manually preprocessed to the lung WS (p=0.023).
Schlussfolgerungen
Detection of COPD with DenseNet models was optimized by manual and automated WSO of the input data to the emphysema WS range, where presence of COPD is highlighted.