Purpose
COVID-19 has infected millions of people worldwide. One of the most important hurdles in controlling the spread of this disease is the inefficiency and lack of medical tests. Computed tomography (CT) scans are promising in providing accurate and fast detection of COVID-19. However, determining COVID-19 requires highly trained radiologists and suffers from inter-observer variability. To remedy these limitations, this paper introduces an automatic methodology based on an ensemble of deep transfer learning for the detection of COVID-19.
Methods
A total of 15 pre-trained convolutional neural networks (CNNs) architectures: EfficientNets(B0-B5), NasNetLarge, NasNetMobile, InceptionV3, ResNet-50, SeResnet 50, Xception, DenseNet121, ResNext50 and Inception_resnet_v2 are used and then fine-tuned on the target task. After that, we built an ensemble method based on majority voting of the best combination of deep transfer learning outputs to further improve the recognition performance. We have used a publicly available dataset of CT scans, which consists of 349 CT scans labeled as being positive for COVID-19 and 397 negative COVID-19 CT scans that are normal or contain other types of lung diseases.
Results
The experimental results indicate that the majority voting of 5 deep transfer learning architecture with EfficientNetB0, EfficientNetB3, EfficientNetB5, Inception_resnet_v2, and Xception has the higher results than the individual transfer learning structure and among the other models based on precision (0.857), recall (0.854) and accuracy (0.85) metrics in diagnosing COVID-19 from CT scans.
Conclusion
Our study based on an ensemble deep transfer learning system with different pre-trained CNNs architectures can work well on a publicly available dataset of CT images for the diagnosis of COVID-19 based on CT scans.
The automatic detection of end-diastole and end-systole frames of echocardiography images is the first step for calculation of the ejection fraction, stroke volume and some other features related to heart motion abnormalities. In this paper, the manifold learning algorithm is applied on 2D echocardiography images to find out the relationship between the frames of one cycle of heart motion. By this approach the nonlinear embedded information in sequential images is represented in a two-dimensional manifold by the LLE algorithm and each image is depicted by a point on reconstructed manifold. There are three dense regions on the manifold which correspond to the three phases of cardiac cycle ('isovolumetric contraction', 'isovolumetric relaxation', 'reduced filling'), wherein there is no prominent change in ventricular volume. By the fact that the end-systolic and end-diastolic frames are in isovolumic phases of the cardiac cycle, the dense regions can be used to find these frames. By calculating the distance between consecutive points in the manifold, the isovolumic frames are mapped on the three minimums of the distance diagrams which were used to select the corresponding images. The minimum correlation between these images leads to detection of end-systole and end-diastole frames. The results on six healthy volunteers have been validated by an experienced echo cardiologist and depict the usefulness of the presented method.
A challenging issue for echocardiographic image interpretation is the accurate analysis of small transient motions of myocardium and valves during real-time visualization. A higher frame rate video may reduce this difficulty, and temporal super resolution (TSR) is useful for illustrating the fast-moving structures. In this paper, we introduce a novel framework that optimizes TSR enhancement of echocardiographic images by utilizing temporal information and sparse representation. The goal of this method is to increase the frame rate of echocardiographic videos, and therefore enable more accurate analyses of moving structures. For the proposed method, we first derived temporal information by extracting intensity variation time curves (IVTCs) assessed for each pixel. We then designed both low-resolution and high-resolution overcomplete dictionaries based on prior knowledge of the temporal signals and a set of prespecified known functions. The IVTCs can then be described as linear combinations of a few prototype atoms in the low-resolution dictionary. We used the Bayesian compressive sensing (BCS) sparse recovery algorithm to find the sparse coefficients of the signals. We extracted the sparse coefficients and the corresponding active atoms in the low-resolution dictionary to construct new sparse coefficients corresponding to the high-resolution dictionary. Using the estimated atoms and the high-resolution dictionary, a new IVTC with more samples was constructed. Finally, by placing the new IVTC signals in the original IVTC positions, we were able to reconstruct the original echocardiography video with more frames. The proposed method does not require training of low-resolution and high-resolution dictionaries, nor does it require motion estimation; it does not blur fast-moving objects, and does not have blocking artifacts.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.