A Two-Stage Deep Learning Strategy for Pneumothorax Classification

Abstract

Due to pneumothorax lesions in chest X-ray images show wide and complicated variation in size, shape, and location within lung regions that overlap with many other anatomic structures such as ribs and vessels, it is a challenge to develop a reliable computer-aided diagnosis systems (CADs) for automatic pneumothorax screening. To address this challenge, we propose a new two-stage deep learning strategy: local feature learning (LFL) followed by global multiple instance learning (GMIL). The GMIL stage intends to train a model that regards the given image as a set of patches and determines whether or not the image contains pneumothorax based on the patches. More specifically, the GMIL model first extracts the hierarchical feature map of a given image by using convolution layer, and takes the feature map of the last layer as a set of depth instances. Each instance is then provided to additional layers to produce its contribution to the final image level prediction. However, the GMIL model trained directly using the original image may still fail to learn highly discriminative features when large areas of non-lesion regions are contained in the image space and thus adversely affect performance. To resolve this problem, prior to the GMIL stage, another model with identical convolutional layers is first trained in the LFL stage using normal patches and pneumothorax-infected patches so that it can better learn the key distinguishing features by reducing most of the non-lesion regions in the X-ray image. The pre-trained convolutional weights are then utilized via transfer learning to enhance training of the GMIL model. Experiments carried on the benchmark ChestX-ray14 data set demonstrate that the proposed learning strategy can achieve the most advanced performance on accuracy, area under receiver operating characteristic curve (AUC), recall, specificity, and F1 scores of 94.4±0. 7%, 97.3±0.5%, 94.6±1.5%, 94.2±0.4% and 94.4±0.7%, separately. We demonstrate the importance and effectiveness of reducing most of the non-lesion regions in the images for learning more discriminative features. The results show that our proposed CAD system is an effective auxiliary tool for screening pneumothorax.