Abstract

Abstract

Abstract
Large-scale datasets are essential for training deep learning models in medical imaging. However, many of these datasets contain poor-quality images that can compromise model performance and clinical reliability. In this study, we propose a framework to detect non-compliant images, such as corrupted scans, incomplete thorax X-rays, and images of non-thoracic body parts, by leveraging contrastive learning for feature extraction and parametric or non-parametric scoring methods for out-of-distribution ranking. Our approach was developed and tested on the CheXpert dataset, achieving an AUC of 0.75 in a manually labeled subset of 1,000 images, and further qualitatively and visually validated on the external PadChest dataset, where it also performed effectively. Our results demonstrate the potential of contrastive learning to detect non-compliant images in large-scale medical datasets, laying the foundation for future work on reducing dataset pollution and improving the robustness of deep learning models in clinical practice. © 2024 IEEE.

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
An automatic system for pathology classification in chest X-ray scans needs more than predictive performance, since providing explanations is deemed essential for fostering end-user trust, improving decision-making, and regulatory compliance. CLARE-XR is a novel methodology that, when presented with an X-ray image, identifies the associated pathologies and provides explanations based on the presentation of similar cases. The diagnosis is achieved using a regression model that maps an image into a 2D latent space containing the reference coordinates of all findings. The references are generated once through label embedding, before the regression step, by converting the original binary ground-truth annotations to 2D coordinates. The classification is inferred minding the distance from the coordinates of an inference image to the reference coordinates. Furthermore, as the regressor is trained on a known set of images, the distance from the coordinates of an inference image to the coordinates of the training set images also allows retrieving similar instances, mimicking the common clinical practice of comparing scans to confirm diagnoses. This inherently interpretable framework discloses specific classification rules and visual explanations through automatic image retrieval methods, outperforming the multi-label ResNet50 classification baseline across multiple evaluation settings on the NIH ChestX-ray14 dataset.

Abstract
The increase in complexity of deep learning models demands explanations that can be obtained with methods like Grad-CAM. This method computes an importance map for the last convolutional layer relative to a specific class, which is then upsampled to match the size of the input. However, this final step assumes that there is a spatial correspondence between the last feature map and the input, which may not be the case. We hypothesize that, for models with large receptive fields, the feature spatial organization is not kept during the forward pass, which may render the explanations devoid of meaning. To test this hypothesis, common architectures were applied to a medical scenario on the public VinDr-CXR dataset, to a subset of ImageNet and to datasets derived from MNIST. The results show a significant dispersion of the spatial information, which goes against the assumption of Grad-CAM, and that explainability maps are affected by this dispersion. Furthermore, we discuss several other caveats regarding Grad-CAM, such as feature map rectification, empty maps and the impact of global average pooling or flatten layers. Altogether, this work addresses some key limitations of Grad-CAM which may go unnoticed for common users, taking one step further in the pursuit for more reliable explainability methods.