Abstract
Backed by more powerful computational resources and optimized training routines, deep learning models have attained unprecedented performance in extracting information from chest X-ray data. Preceding other tasks, an automated abnormality detection stage can be useful to prioritize certain exams and enable a more efficient clinical workflow. However, the presence of image artifacts such as lettering often generates a harmful bias in the classifier, leading to an increase of false positive results. Consequently, healthcare would benefit from a system that selects the thoracic region of interest prior to deciding whether an image is possibly pathologic. The current work tackles this binary classification exercise using an attention-driven and spatially unsupervised Spatial Transformer Network (STN). The results indicate that the STN achieves similar results to using YOLO-cropped images, with fewer computational expenses and without the need for localization labels. More specifically, the system is able to distinguish between normal and abnormal CheXpert images with a mean AUC of 84.22%.

Abstract
Abstract Funding Acknowledgements: Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Health Research Council of New Zealand (HRC) National Heart Foundation of New Zealand (NHF) Segmentation of the left ventricular myocardium and cavity in 3D echocardiography (3DE) is a critical task for the quantification of systolic function in heart disease. Continuing advances in 3DE have considerably improved image quality, prompting increased clinical uptake in recent years, particularly for volumetric measurements. Nevertheless, analysis of 3DE remains a difficult problem due to inherently complex noise characteristics, anisotropic image resolution, and regions of acoustic dropout. One of the primary challenges associated with the development of automated methods for 3DE analysis is the requirement of a sufficiently large training dataset. Historically, ground truth annotations have been difficult to obtain due to the high degree of inter- and intra-observer variability associated with manual 3DE segmentation, thus, limiting the scope of AI-based solutions. To address the lack of expert consensus, we instead used labels derived from cardiac magnetic resonance (CMR) images of the same subjects. By spatiotemporally registering CMR labels to corresponding 3DE image data on a per subject basis (Figure 1), we collated 520 annotated 3DE images from a mixed cohort of 130 human subjects (2 independent single-beat acquisitions per subject at end-diastole and end-systole) consisting of healthy controls and patients with acquired cardiac disease. Comprising images acquired across a range of patient demographics, this curated dataset exhibits variation in image quality, 3DE acquisition parameters, as well as left ventricular shape and pose within the 3D image volume. To demonstrate the utility of such a dataset, nn-UNet, a self-configuring deep learning method for semantic segmentation was employed. An 80/20 split of the dataset was used for training and testing, respectively, and data augmentations were applied in the form of scaling, rotation, and reflection. The trained network was capable of reproducing measurements derived from CMR for end-diastolic volume, end-systolic volume, ejection fraction, and mass, while outperforming an expert human observer in terms of accuracy as well as scan-rescan reproducibility (Table I). As part of ongoing efforts to improve the accuracy and efficiency of 3DE analysis, we have leveraged the high resolution and signal-to-noise-ratio of CMR (relative to 3DE), to create a novel, publicly available benchmark dataset for developing and evaluating 3DE labelling methods. This approach not only significantly reduces the effects of observer-specific bias and variability in training data arising from conventional manual 3DE analysis methods, but also improves the agreement between cardiac indices derived from 3DE and CMR. Figure 1. Data annotation workflow Table I. Results

Abstract
The coronavirus disease 2019 (COVID-19) evolved into a global pandemic, responsible for a significant number of infections and deaths. In this scenario, point-of-care ultrasound (POCUS) has emerged as a viable and safe imaging modality. Computer vision (CV) solutions have been proposed to aid clinicians in POCUS image interpretation, namely detection/segmentation of structures and image/patient classification but relevant challenges still remain. As such, the aim of this study is to develop CV algorithms, using Deep Learning techniques, to create tools that can aid doctors in the diagnosis of viral and bacterial pneumonia (VP and BP) through POCUS exams. To do so, convolutional neural networks were designed to perform in classification tasks. The architectures chosen to build these models were the VGG16, ResNet50, DenseNet169 e MobileNetV2. Patients images were divided in three classes: healthy (HE), BP and VP (which includes COVID-19). Through a comparative study, which was based on several performance metrics, the model based on the DenseNet169 architecture was designated as the best performing model, achieving 78% average accuracy value of the five iterations of 5- Fold Cross-Validation. Given that the currently available POCUS datasets for COVID-19 are still limited, the training of the models was negatively affected by such and the models were not tested in an independent dataset. Furthermore, it was also not possible to perform lesion detection tasks. Nonetheless, in order to provide explainability and understanding of the models, Gradient-weighted Class Activation Mapping (GradCAM) were used as a tool to highlight the most relevant classification regions. Clinical relevance - Reveals the potential of POCUS to support COVID-19 screening. The results are very promising although the dataset is limite

Abstract
Echocardiography-based cardiac boundary tracking provides valuable information about the heart condition for interventional procedures and intensive care applications. Nevertheless, echocardiographic images come with several issues, making it a challenging task to develop a tracking and segmentation algorithm that is robust to shadows, occlusions, and heart rate changes. We propose an autonomous tracking method to improve the robustness and efficiency of echocardiographic tracking. A method denoted by hybrid Condensation and adaptive Kalman filter (HCAKF) is proposed to overcome tracking challenges of echocardiograms, such as variable heart rate and sensitivity to the initialization stage. The tracking process is initiated by utilizing active shape model, which provides the tracking methods with a number of tracking features. The procedure tracks the endocardium borders, and it is able to adapt to changes in the cardiac boundaries velocity and visibility. HCAKF enables one to use a much smaller number of samples that is used in Condensation without sacrificing tracking accuracy. Furthermore, despite combining the two methods, our complexity analysis shows that HCAKF can produce results in real-time. The obtained results demonstrate the robustness of the proposed method to the changes in the heart rate, yielding an Hausdorff distance of 1.032 +/- 0.375 while providing adequate efficiency for real-time operations.

Abstract

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.