Abstract
Chest radiography (CXR) is one of the most used medical imaging modalities. Nevertheless, the interpretation of CXR images is time-consuming and subject to variability. As such, automated systems for pathology detection have been proposed and promising results have been obtained, particularly using deep learning. However, these tools suffer from poor explainability, which represents a major hurdle for their adoption in clinical practice. One proposed explainability method in CXR is through contrastive examples, i.e. by showing an alternative version of the CXR except without the lesion being investigated. While image-level normal/healthy image synthesis has been explored in literature, normal patch synthesis via inpainting has received little attention. In this work, a method to synthesize contrastive examples in CXR based on local synthesis of normal CXR patches is proposed. Based on a contextual attention inpainting network (CAttNet), an anatomically-guided inpainting network (AnaCAttNet) is proposed that leverages anatomical information of the original CXR through segmentation to guide the inpainting for a more realistic reconstruction. A quantitative evaluation of the inpainting is performed, showing that AnaCAttNet outperforms CAttNet (FID of 0.0125 and 0.0132 respectively). Qualitative evaluation by three readers also showed that AnaCAttNet delivers superior reconstruction quality and anatomical realism. In conclusion, the proposed anatomical segmentation module for inpainting is shown to improve inpainting performance.

Abstract

Abstract

Abstract
Deep learning-based models are widely used for disease classification in chest radiographs. This exam can be performed in one of two projections (posteroanterior or anteroposterior), depending on the direction that the X-ray beam travels through the body. Since projection visibly affects the way anatomical structures appear in the scans, it may introduce bias in classifiers, especially when spurious correlations between a given disease and a projection occur. This paper examines the influence of chest radiograph projection on the performance of deep learning-based classification models and proposes an approach to mitigate projection-induced bias. Results show that a DenseNet-121 model is better at classifying images from the most representative projection in the data set, suggesting that projection is taken into account by the classifier. Moreover, this model can classify chest X-ray projection better than any of the fourteen radiological findings considered, without being explicitly trained for that task, putting it at high risk for projection bias. We propose a label-conditional gradient reversal framework to make the model insensitive to projection, by forcing the extracted features to be simultaneously good for disease classification and bad for projection classification, resulting in a framework with reduced projection-induced bias.

Abstract

Abstract