Big Self-Supervised Models Advance Medical Image Classification

2021 IEEE/CVF International Conference on Computer Vision (ICCV), Journal Year: 2021, Volume and Issue: unknown, P. 3458 - 3468

Published: Oct. 1, 2021

Self-supervised pretraining followed by supervised fine-tuning has seen success in image recognition, especially when labeled examples are scarce, but received limited attention medical analysis. This paper studies the effectiveness of self-supervised learning as a pre-training strategy for classification. We conduct experiments on two distinct tasks: dermatology condition classification from digital camera images and multi-label chest X-ray classification, demonstrate that ImageNet, additional unlabeled domain-specific significantly improves accuracy classifiers. introduce novel Multi-Instance Contrastive Learning (MICLe) method uses multiple underlying pathology per patient case, available, to construct more informative positive pairs learning. Combining our contributions, we achieve an improvement 6.7% top-1 1.1% mean AUC respectively, outperforming strong baselines pretrained ImageNet. In addition, show big models robust distribution shift can learn efficiently with small number images.

Language: Английский

---

Tools and Techniques for Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2)/COVID-19 Detection

Clinical Microbiology Reviews, Journal Year: 2021, Volume and Issue: 34(3)

Published: May 11, 2021

The coronavirus disease 2019 (COVID-19) pandemic, caused by severe acute respiratory 2 (SARS-CoV-2), has led to millions of confirmed cases and deaths worldwide. Efficient diagnostic tools are in high demand, as rapid large-scale testing plays a pivotal role patient management decelerating spread.

Language: Английский

---

Deep transfer learning-based automated detection of COVID-19 from lung CT scan slices

Applied Intelligence, Journal Year: 2020, Volume and Issue: 51(1), P. 571 - 585

Published: Aug. 21, 2020

Language: Английский

---

A Review on Deep Learning Techniques for the Diagnosis of Novel Coronavirus (COVID-19)

IEEE Access, Journal Year: 2021, Volume and Issue: 9, P. 30551 - 30572

Published: Jan. 1, 2021

Novel coronavirus (COVID-19) outbreak, has raised a calamitous situation all over the world and become one of most acute severe ailments in past hundred years. The prevalence rate COVID-19 is rapidly rising every day throughout globe. Although no vaccines for this pandemic have been discovered yet, deep learning techniques proved themselves to be powerful tool arsenal used by clinicians automatic diagnosis COVID-19. This paper aims overview recently developed systems based on using different medical imaging modalities like Computer Tomography (CT) X-ray. review specifically discusses provides insights well-known data sets train these networks. It also highlights partitioning various performance measures researchers field. A taxonomy drawn categorize recent works proper insight. Finally, we conclude addressing challenges associated with use methods detection probable future trends research area. aim facilitate experts (medical or otherwise) technicians understanding ways are regard how they can potentially further utilized combat outbreak

Language: Английский

---

COVID-19 detection in CT images with deep learning: A voting-based scheme and cross-datasets analysis

Informatics in Medicine Unlocked, Journal Year: 2020, Volume and Issue: 20, P. 100427 - 100427

Published: Jan. 1, 2020

Early detection and diagnosis are critical factors to control the COVID-19 spreading. A number of deep learning-based methodologies have been recently proposed for screening in CT scans as a tool automate help with diagnosis. These approaches, however, suffer from at least one following problems: (i) they treat each scan slice independently (ii) methods trained tested sets images same dataset. Treating slices means that patient may appear training test time which produce misleading results. It also raises question whether should be evaluated group or not. Moreover, using single dataset concerns about generalization methods. Different datasets tend present varying quality come different types machines reflecting conditions countries cities where from. In order address these two problems, this work, we propose an Efficient Deep Learning Technique voting-based approach. approach, given classified voting system. The approach is biggest analysis patient-based split. cross study presented assess robustness models more realistic scenario data comes distributions. cross-dataset has shown power learning far acceptable task since accuracy drops 87.68% 56.16% on best evaluation scenario. results highlighted aim CT-images improve significantly considered clinical option larger diverse needed evaluate

Language: Английский

---

Contrastive Cross-Site Learning With Redesigned Net for COVID-19 CT Classification

IEEE Journal of Biomedical and Health Informatics, Journal Year: 2020, Volume and Issue: 24(10), P. 2806 - 2813

Published: Sept. 10, 2020

The pandemic of coronavirus disease 2019 (COVID-19) has lead to a global public health crisis spreading hundreds countries. With the continuous growth new infections, developing automated tools for COVID-19 identification with CT image is highly desired assist clinical diagnosis and reduce tedious workload interpretation. To enlarge datasets machine learning methods, it essentially helpful aggregate cases from different medical systems robust generalizable models. This paper proposes novel joint framework perform accurate by effectively heterogeneous distribution discrepancy. We build powerful backbone redesigning recently proposed COVID-Net in aspects network architecture strategy improve prediction accuracy efficiency. On top our improved backbone, we further explicitly tackle cross-site domain shift conducting separate feature normalization latent space. Moreover, propose use contrastive training objective enhance invariance semantic embeddings boosting classification performance on each dataset. develop evaluate method two large-scale made up images. Extensive experiments show that approach consistently improves performanceson both datasets, outperforming original trained dataset 12.16% 14.23% AUC respectively, also exceeding existing state-of-the-art multi-site methods.

Language: Английский

---

Explainable COVID-19 Detection Using Chest CT Scans and Deep Learning

Sensors, Journal Year: 2021, Volume and Issue: 21(2), P. 455 - 455

Published: Jan. 11, 2021

This paper explores how well deep learning models trained on chest CT images can diagnose COVID-19 infected people in a fast and automated process. To this end, we adopted advanced network architectures proposed transfer strategy using custom-sized input tailored for each architecture to achieve the best performance. We conducted extensive sets of experiments two image datasets, namely, SARS-CoV-2 CT-scan COVID19-CT. The results show superior performances our compared with previous studies. Our achieved average accuracy, precision, sensitivity, specificity, F1-score values 99.4%, 99.6%, 99.8%, 99.4% dataset, 92.9%, 91.3%, 93.7%, 92.2%, 92.5% COVID19-CT respectively. For better interpretability results, applied visualization techniques provide visual explanations models’ predictions. Feature visualizations learned features well-separated clusters representing non-COVID-19 cases. Moreover, indicate that are not only capable identifying cases but also accurate localization COVID-19-associated regions, as indicated by well-trained radiologists.

Language: Английский

---

Applications of artificial intelligence in battling against covid-19: A literature review

Chaos Solitons & Fractals, Journal Year: 2020, Volume and Issue: 142, P. 110338 - 110338

Published: Oct. 3, 2020

Language: Английский

---

MetaCOVID: A Siamese neural network framework with contrastive loss for n-shot diagnosis of COVID-19 patients

Pattern Recognition, Journal Year: 2020, Volume and Issue: 113, P. 107700 - 107700

Published: Oct. 17, 2020

Language: Английский

---

CERT: Contrastive Self-supervised Learning for Language Understanding

Published: May 21, 2020

Pretrained language models such as BERT, GPT have shown great effectiveness in understanding. The auxiliary predictive tasks existing pretraining approaches are mostly defined on tokens, thus may not be able to capture sentence-level semantics very well. To address this issue, we propose CERT: Contrastive self-supervised Encoder Representations from Transformers, which pretrains representation using contrastive learning at the sentence level. CERT creates augmentations of original sentences back-translation. Then it finetunes a pretrained encoder (e.g., BERT) by predicting whether two augmented originate same sentence. is simple use and can flexibly plugged into any pretraining-finetuning NLP pipeline. We evaluate three understanding tasks: CoLA, RTE, QNLI. outperforms BERT significantly.<br>

Language: Английский

---