Background: Deep learning is the state-of-the-art approach for automated segmentation of the left ventricle (LV) and right ventricle (RV) in cardiovascular magnetic resonance (CMR) images. However, these models have been mostly trained and validated using CMR datasets of structurally normal hearts or cases with acquired cardiac disease, and are therefore not well-suited to handle cases with congenital cardiac disease such as tetralogy of Fallot (TOF). We aimed to develop and validate a dedicated model with improved performance for LV and RV cavity and myocardium quantification in patients with repaired TOF. Methods: We trained a three-dimensional (3D) convolutional neural network (CNN) with 5-fold cross-validation using manually delineated end-diastolic (ED) and end-systolic (ES) short-axis image stacks obtained from either a public dataset containing patients with no or acquired cardiac pathology (n = 100), an institutional dataset of TOF patients (n = 96), or both datasets mixed. Our method allows for missing labels in the training images to accommodate for different ED and ES phases for LV and RV as is commonly the case in TOF. The best performing model was applied to all frames of a separate test set of TOF cases (n = 36) and ED and ES phases were automatically determined for LV and RV separately. The model was evaluated against the performance of a commercial software (suiteHEART (R), NeoSoft, Pewaukee, Wisconsin, US). Results: Training on the mixture of both datasets yielded the best agreement with the manual ground truth for the TOF cases, achieving a median Dice similarity coefficient of (93.8%, 89.8%) for LV cavity and of (92.9%, 90.9%) for RV cavity at (ED, ES) respectively, and of 80.9% and 61.8% for LV and RV myocardium at ED. The offset in automated ED and ES frame selection was 0.56 and 0.89 frames on average for LV and RV respectively. No statistically significant differences were found between our model and the commercial software for LV quantification (two-sided Wilcoxon signed rank test, p < 5%), while RV quantification was significantly improved with our model achieving a mean absolute error of 12 ml for RV cavity compared to 36 ml for the commercial software. Conclusion: We developed and validated a fully automatic segmentation and quantification approach for LV and RV, including RV mass, in patients with repaired TOF. Compared to a commercial software, our approach is superior for RV quantification indicating its potential in clinical practice.

Albeit the Dice loss is one of the dominant loss functions in medical image segmentation, most research omits a closer look at its derivative, i.e. the real motor of the optimization when using gradient descent. In this paper, we highlight the peculiar action of the Dice loss in the presence of missing or empty labels. First, we formulate a theoretical basis that gives a general description of the Dice loss and its derivative. It turns out that the choice of the reduction dimensions and the smoothing term ϵ is non-trivial and greatly influences its behavior. We find and propose heuristic combinations of and ϵ that work in a segmentation setting with either missing or empty labels. Second, we empirically validate these findings in a binary and multiclass segmentation setting using two publicly available datasets. We confirm that the choice of and ϵ is indeed pivotal. With chosen such that the reductions happen over a single batch (and class) element and with a negligible ϵ , the Dice loss deals with missing labels naturally and performs similarly compared to recent adaptations specific for missing labels. With chosen such that the reductions happen over multiple batch elements or with a heuristic value for ϵ , the Dice loss handles empty labels correctly. We believe that this work highlights some essential perspectives and hope that it encourages researchers to better describe their exact implementation of the Dice loss in future work.

Semantic segmentation using convolutional neural networks (CNNs) is the state-of-theart for many medical image segmentation tasks including myocardial segmentation in cardiac MR images.However, the predicted segmentation maps obtained from such standard CNN do not allow direct quantification of regional shape properties such as regional wall thickness.Furthermore, the CNNs lack explicit shape constraints, occasionally resulting in unrealistic segmentations.In this paper, we use a CNN to predict shape parameters of an underlying statistical shape model of the myocardium learned from a training set of images.Additionally, the cardiac pose is predicted, which allows to reconstruct the myocardial contours.The integrated shape model regularizes the predicted contours and guarantees realistic shapes.We enforce robustness of shape and pose prediction by simultaneously performing pixel-wise semantic segmentation during training and define two loss functions to impose consistency between the two predicted representations: one distance-based loss and one overlap-based loss.We evaluated the proposed method in a 5-fold cross validation on an in-house clinical dataset with 75 subjects and on the ACDC and LVQuan19 public datasets.We show that the two newly defined loss functions successfully increase the consistency between shape and pose parameters and semantic segmentation, which leads to a significant improvement of the reconstructed myocardial contours.Additionally, these loss functions drastically reduce the occurrence of unrealistic shapes in the semantic segmentation output.

Semantic segmentation using convolutional neural networks (CNNs) is the state-of-the-art for many medical segmentation tasks including left ventricle (LV) segmentation in cardiac MR images. However, a drawback is that these CNNs lack explicit shape constraints, occasionally resulting in unrealistic segmentations. In this paper, we perform LV and myocardial segmentation by regression of pose and shape parameters derived from a statistical shape model. The integrated shape model regularizes predicted segmentations and guarantees realistic shapes. Furthermore, in contrast to semantic segmentation, it allows direct calculation of regional measures such as myocardial thickness. We enforce robustness of shape and pose prediction by simultaneously constructing a segmentation distance map during training. We evaluated the proposed method in a fivefold cross validation on a in-house clinical dataset with 75 subjects containing a total of 1539 delineated short-axis slices covering LV from apex to base, and achieved a correlation of 99% for LV area, 94% for myocardial area, 98% for LV dimensions and 88% for regional wall thicknesses. The method was additionally validated on the LVQuan18 and LVQuan19 public datasets and achieved state-of-the-art results.

Abstract Funding Acknowledgements Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Ph.D fellowship of the Research Foundation Flanders (FWO). The Master@Heart trial is funded by the FWO. Introduction Differentiating intensive training induced hypertrophy from hyperthropic cardiomyopathy (HCM) is important to identify those young athletes at risk of sudden cardiac death. Swoboda and colleagues demonstrated that T1 and ECV mapping can aid such a differentiation between athletic and pathological hypertrophy, particularly in subjects with indeterminate wall thickness (1). Recently texture analysis (TA) methods of CMR data have demonstrated improved diagnostic accuracy over conventional qualitative analysis in various heart diseases. Only few studies have applied TA to T1 and ECV mapping data (2-4). Here we aimed to demonstrate that a TA approach provides superior capacity to distinguish HCM from athlete’s heart over average native T1 and ECV values. Purpose It was our hypothesis that a texture analysis of T1 and ECV mapping images would identify features that could discriminate between a HCM and athlete’s heart with a higher classification accuracy (CA) than average T1 and ECV values. Methods This study included data from 97 subjects diagnosed with HCM (acc. to guidelines; 5) and 28 athletes that took part in the Master@Heart trial (an ongoing study assessing the beneficial effects of long-term endurance exercise for the prevention of coronary artery disease, 6). Long and short axis T1 mapping data was acquired on a 1.5T Philips Ingenia system using MOLLI (seconds scheme). After offline motion correction and T1 and ECV map calculation (7), the left ventricular myocardium was manually delineated (3D Slicer; 8). Texture analysis of the masked images resulted in 194 features (Pyradiomics, standard settings; 9). The dataset was then split (75/25%) for training and testing purposes keeping images from the same subject within the same set. A fast correlation based filter rank was applied to the training data to derive relevant features. A further reduction to only two features was based on the CA of a support vector machine (SVM) learning method (linear kernel; cost 0.9 regression loss epsilon 0.1; leave-one-out). Finally, ROC analysis on the test data was used to determine the diagnostic accuracy for the following predictors: (1) median T1 and ECV (2) two most relevant features (training) (3) combination of (1) and (2) (ROC AUC statistics (10)). Results The two most relevant features were the histogram feature ECV energy and the gray level size zone matrix (GLSZM) feature native T1 zone entropy, a measure of heterogeneity in the texture pattern. A model to distinguish HCM from athletes based on these features outperformed the model using only median T1 and ECV values with both higher sensitivity and specificity (table 1) and a significantly higher AUC in the ROC analysis (p < 0.05, figure 1). Combining these two features with median values did not improve the CA further. Conclusion Texture analysis of motion-corrected T1 and ECV mapping images out-performs classical analysis based on average values in distinguishing HCM from athlete"s heart.

Quantification of left ventricular (LV) parameters from cardiac MRI is important to assess cardiac condition and help in the diagnosis of certain pathologies. We present a CNN-based approach for automatic quantification of 11 LV indices: LV and myocardial area, 3 LV dimensions and 6 regional wall thicknesses (RWT). We use an encoder-decoder segmentation architecture and hypothesize that deep feature maps contain important shape information suitable to start an additional network branch for LV index regression. The CNN is simultaneously trained on regression and segmentation losses. We validated our approach on the LVQuan19 training dataset and found that our proposed CNN significantly outperforms a standard encoder regression CNN. The mean absolute error and Pearson correlation coefficient obtained for the different indices are respectively 190mm2 (96%), 214mm2 (0.90%), 2.99mm (95%) and 1.82mm (71%) for LV area, myocardial area, LV dimensions and RWT on a three-fold cross validation and 186mm(2) (97%), 222mm(2) (0.88%), 3.03mm (0.95%) and 1.67mm (73%) on a five-fold cross validation.

Recent research on COVID-19 suggests that CT imaging provides useful information to assess disease progression and assist diagnosis, in addition to help understanding the disease. There is an increasing number of studies that propose to use deep learning to provide fast and accurate quantification of COVID-19 using chest CT scans. The main tasks of interest are the automatic segmentation of lung and lung lesions in chest CT scans of confirmed or suspected COVID-19 patients. In this study, we compare twelve deep learning algorithms using a multi-center dataset, including both open-source and in-house developed algorithms. Results show that ensembling different methods can boost the overall test set performance for lung segmentation, binary lesion segmentation and multiclass lesion segmentation, resulting in mean Dice scores of 0.982, 0.724 and 0.469, respectively. The resulting binary lesions were segmented with a mean absolute volume error of 91.3 ml. In general, the task of distinguishing different lesion types was more difficult, with a mean absolute volume difference of 152 ml and mean Dice scores of 0.369 and 0.523 for consolidation and ground glass opacity, respectively. All methods perform binary lesion segmentation with an average volume error that is better than visual assessment by human raters, suggesting these methods are mature enough for a large-scale evaluation for use in clinical practice.

Accurate left ventricular (LV) segmentation in cardiac MRI facilitates quantification of clinical parameters such as LV volume and ejection fraction (EF). We present a CNN-based method to obtain a 3D representation of LV by integrating information from 2D short-axis and horizontal and vertical long-axis images. Our CNN is flexible to the number of input slices and uses an additional input of image coordinates as spatial context. This concept is validated on variations of two well-known CNN architectures for medical image segmentation: U-Net and DeepMedic. Five-fold cross validation on a dataset of 20 patients achieved a correlation of 95.0/93.1% for quantification of end-diastolic volume, 91.6/90.8% for end-systolic volume and 80.5/84.5% for EF for the two architectures respectively. We show that (1) incorporating long-axis data improves segmentation performance and (2) providing spatial context by adding image coordinates as input to the CNN yields similar performance with a smaller receptive field.

Our motivating application is a real-world problem: COVID-19 classification from CT imaging, for which we present an explainable Deep Learning approach based on a semi-supervised classification pipeline that employs variational autoencoders to extract efficient feature embedding. We have optimized the architecture of two different networks for CT images: (i) a novel conditional variational autoencoder (CVAE) with a specific architecture that integrates the class labels inside the encoder layers and uses side information with shared attention layers for the encoder, which make the most of the contextual clues for representation learning, and (ii) a downstream convolutional neural network for supervised classification using the encoder structure of the CVAE. With the explainable classification results, the proposed diagnosis system is very effective for COVID-19 classification. Based on the promising results obtained qualitatively and quantitatively, we envisage a wide deployment of our developed technique in large-scale clinical studies.Code is available at https://git.etrovub.be/AVSP/ct-based-covid-19-diagnostic-tool.git.

T1 and ECV mapping are quantitative methods for myocardial tissue characterization using cardiac MRI, and are highly relevant for the diagnosis of diffuse myocardial diseases. Since the maps are calculated pixel-by-pixel from a set of MRI images with different T1-weighting, it is critical to assure exact spatial correspondence between these images. However, in practice, different sources of motion e.g. cardiac motion, respiratory motion or patient motion, hamper accurate T1 and ECV calculation such that retrospective motion correction is required. We propose a new robust non-rigid registration framework combining a data-driven initialization with a model-based registration approach, which uses a model for T1 relaxation to avoid direct registration of images with highly varying contrast. The registration between native T1 and enhanced T1 to obtain a motion free ECV map is also calculated using information from T1 model fitting. The method was validated on three datasets recorded with two substantially different acquisition protocols (MOLLI (dataset 1 (n=15) and dataset 2 (n=29)) and STONE (dataset 3 (n = 210))), one in breath-hold condition and one free-breathing. The average Dice coefficient increased from 72.6 +/- 12.1% to 82.3 +/- 7.4% (P < 0.05) and mean boundary error decreased from 2.91 +/- 1.51mm to 1.62 +/- 0.80mm (P < 0.05) for motion correction in a single T1-weighted image sequence (3 datasets) while average Dice coefficient increased from 63.4 +/- 22.5% to 79.2 +/- 8.5% (P < 0.05) and mean boundary error decreased from 3.26 +/- 2.64mm to 1.77 +/- 0.86mm (P < 0.05) between native and enhanced sequences (dataset 1 and 2). Overall, the native T1 SD error decreased from 67.32 +/- 32.57ms to 58.11 +/- 21.59ms (P < 0.05), enhanced SD error from 30.15 +/- 25ms to 22.74 +/- 8.94ms (P < 0.05) and ECV SD error from 10.08 +/- 9.59% to 5.42 +/- 3.21% (P <0.05) (dataset 1 and 2). (C) 2018 Elsevier B.V. All rights reserved.

Quantification of myocardial T1 and extra cellular volume (ECV) in cardiac MRI provides relevant diagnostic information about myocardial structure. However, since these maps are pixel-wise calculated from two sequences of T1-weighted images, they are frequently disturbed by motion artifacts originating from e.g. patient motion, failure in breath-hold or cardiac motion. We propose a new non-rigid registration framework combining a robust data-driven initialization with a model-based refinement. The data-driven algorithm finds an optimal registration sequence of images for calculation of an initial T1 map. The registration is subsequently refined by exploiting the exponential relaxation model of T1. Validation using 20 in-vivo data sets showed a decrease in mean boundary error and an increase in global Dice coefficient.

Acute myocardial infarction is a complex clinical condition that affects the myocardium in multiple ways. Cardiac magnetic resonance provides a range of noninvasive sequences that can optimally characterize jeopardized myocardium in acute coronary syndrome patients through both visual and

Several studies suggest that radiomics of non-contrast cine balanced steady-state free precession (bSSFP) images allows to detect myocardial infarct patients. Relevant radiomic features vary however without clear knowledge of the underlying causes. Most studies rely on time consuming manual contouring on use end-diastolic images while it was shown that features vary through the heart cycle. In this study we show that radiomics of end-systolic images result in a higher classification accuracy than for end-diastolic images and produce the same result when relying on commercial automated contouring.

Pre-contrast balanced steady-state free precession (bssfp) cine images seem to contain sufficient information to identify patients with an acute or chronic myocardial infarction using radiomics or deep learning methods. The reason remains however to be elucidated. It is also unclear if the cardiac phase should be considered although radiomics features vary through the heart cycle. In this study we show that the known global myocardial bssfp signal intensity variations observed in healthy subject are in fact segment dependent and have a reproducible three peak profile. In patients with acute myocardial infarct these profiles shows segment dependent alterations vs. controls.