Source-linked AI summary
Automated Detection and Forecasting of COVID-19 using Deep Learning Techniques: A Review
Afshin Shoeibi, Marjane Khodatars, Mahboobeh Jafari, Navid Ghassemi, Delaram Sadeghi, Parisa Moridian, Ali Khadem, Roohallah Alizadehsani, Sadiq Hussain, Assef Zare, Zahra Alizadeh Sani, Fahime Khozeimeh, Saeid Nahavandi, U. Rajendra Acharya, Juan M. Gorriz
TL;DR
COVID-19 imaging diagnosis is challenging because it is time-consuming and prone to human error, while forecasting is also constrained by limited data and disease complexity. This paper surveys deep-learning studies using X-ray and CT images for detection and lung segmentation, reviews prevalence forecasting, and discusses challenges and future directions. The review reports broad use of these approaches, including a cited accuracy of 99.26% in one study, while identifying data availability and related limitations as central challenges.
Problem
COVID-19 detection from X-ray and CT images is time-consuming and prone to human errors, creating a need to review automated approaches.
Method
The paper conducts a comprehensive review of deep-learning studies on COVID-19 detection, lung segmentation, and prevalence forecasting.
Results
The reviewed literature includes detection, segmentation, and forecasting studies, with one cited detection result reporting Acc=99.26%.
Takeaways & Limitations
The review identifies deep-learning applications across COVID-19 imaging diagnosis, lung segmentation, and forecasting, and summarizes challenges and future research directions.
Takeaways & Limitations
Limited public X-ray and CT databases constrain deep-learning development, and forecasting studies additionally require large databases.
Abstract
from arXiv · showhide
Coronavirus, or COVID-19, is a hazardous disease that has endangered the health of many people around the world by directly affecting the lungs. COVID-19 is a medium-sized, coated virus with a single-stranded RNA, and also has one of the largest RNA genomes and is approximately 120 nm. The X-Ray and computed tomography (CT) imaging modalities are widely used to obtain a fast and accurate medical diagnosis. Identifying COVID-19 from these medical images is extremely challenging as it is time-consuming and prone to human errors. Hence, artificial intelligence (AI) methodologies can be used to obtain consistent high performance. Among the AI methods, deep learning (DL) networks have gained popularity recently compared to conventional machine learning (ML). Unlike ML, all stages of feature extraction, feature selection, and classification are accomplished automatically in DL models. In this paper, a complete survey of studies on the application of DL techniques for COVID-19 diagnostic and segmentation of lungs is discussed, concentrating on works that used X-Ray and CT images. Additionally, a review of papers on the forecasting of coronavirus prevalence in different parts of the world with DL is presented. Lastly, the challenges faced in the detection of COVID-19 using DL techniques and directions for future research are discussed.
I. INTRODUCTION
COVID-19 affects the lungs and is difficult to diagnose quickly and reliably with medical imaging. The paper reviews deep-learning approaches for detection, segmentation, and forecasting, along with their challenges.
- COVID-19 can damage lung cells and is difficult to diagnose and treat because of its mutation characteristics.
- RT-PCR is the diagnostic gold standard, but imaging is used for primary detection because RT-PCR is time-consuming.
- Radiologists may find X-ray interpretation prolonged and may mistake diseases such as pulmonary tuberculosis for COVID-19.
- CT provides higher contrast than X-ray, but analyzing numerous patient slices is challenging.
- The review covers deep-learning methods for COVID-19 detection, lung segmentation, and prevalence forecasting.
II. SEARCH STRATEGY
The paper searches multiple scholarly databases and organizes prior work around deep-learning methods for COVID-19 detection, segmentation, and forecasting. It draws on public image and forecasting datasets.
- The literature search used IEEE Xplore, ScienceDirect, SpringerLink, ACM, ArXiv, and Google Scholar through September 19, 2021.
- Search terms covered COVID-19, deep learning, segmentation, forecasting, attention, transformers, data fusion, and graph learning.
- The reviewed computer-aided systems are categorized into classification for identifying patients and segmentation for identifying infected regions.
- Public databases include X-ray and CT datasets for detection and segmentation and datasets for predicting spread in leading countries.
- The review discusses established classification and segmentation architectures, forecasting models, and state-of-the-art methods.
1) Classification Models:
This section presents convolutional and pre-trained deep-learning models as responses to limited data, overfitting risk, and the design effort required for CNNs.
- CNNs accept 2D images with minimal preprocessing and retain structural information among neighboring pixels or voxels.
- Deep models risk overfitting because training samples are limited relative to the number of learnable parameters.
- Researchers use pre-trained architectures because CNN training requires substantial data and structure design is time-consuming.
- Common pre-trained choices include AlexNet, VGG, GoogLeNet, ResNet, DenseNet, and SqueezeNet.
2) Generative Adversarial Networks (GAN):
Generative adversarial networks are reviewed as a data-augmentation strategy for deep learning when medical-image datasets are limited.
- GANs address limited dataset size by generating high-quality synthetic data for augmentation.
- A GAN trains two networks in a minimax game: one distinguishes real from generated data, while the other produces indistinguishable images.
3) Segmentation Models:
The review covers deep-learning models for lung-region segmentation in COVID-19 images, including FCN, SegNet, U-Net, and Res2Net. U-Net uses trainable deconvolution layers and skip connections for segmentation.
- FCN, SegNet, U-Net, and Res2Net are among the deep-learning models developed for lung-region segmentation.
- Figure 6 presents a typical U-Net architecture for segmenting lungs in COVID-19 patients.
- Segmentation networks use an encoder to produce low-resolution maps and a decoder to restore the original image resolution.In SegNet, the decoder places an up-sampling layer after each encoding down-sampling layer.
- U-Net has matching numbers of pooling and up-sampling layers, with trainable deconvolution layers and skip connections between corresponding levels.
4) Forecasting Models:
The review describes recurrent neural networks as models for sequential and time-series data, with LSTM and GRU architectures designed to manage long-term dependencies through gating mechanisms.
- Recurrent neural networks extend feed-forward networks to capture long-term dependencies and features in sequential and time-series data.
- LSTM uses memory, forget, input, and output gates to decide which information is retained, discarded, and organized.
- Unlike vanilla RNNs, LSTMs can preserve potential long-term dependencies through gated memory control.
- GRU combines the forget and input gates into an update gate and merges memory-cell and hidden states into one state.Its update gate controls information addition or discarding, while its reset gate controls how much earlier information is forgotten.
- The gating design makes GRU simpler than LSTM.
DEEP ATTENTION LEARNING
The review covers attention, transformer, fusion, and graph-based deep-learning approaches for COVID-19 diagnosis and related imaging tasks. These methods apply different mechanisms to focus inputs, transform representations, combine features, or model feature relationships.
- DEEP ATTENTION LEARNING: Attention-based models concentrate on subsets of inputs containing task-relevant information and have been used for COVID-19 classification, segmentation, and diagnosis.One reviewed model uses an attention module with VGG-16 to classify COVID-19 from X-ray images.
- DEEP TRANSFORMER LEARNING: Transformer approaches include spatial, graph, recurrent spatial, and Polar transformer networks; ViT has been used to diagnose COVID-19 from ultrasound data.
- DEEP FUSION TECHNIQUES: A deep-fusion approach extracts CT features with four CNNs, integrates and ranks those features, and then applies an SVM classifier.
- GRAPH DEEP LEARNING: A graph deep-learning approach extracts CT features with a 3D-CNN, constructs a COVID-19 graph in a GCN, and combines three deep-learning models for detection.
V. DISCUSSION
The discussion reviews deep-learning studies across COVID-19 detection, lung segmentation, and prevalence forecasting using varied X-ray, CT, ultrasound, and combined datasets. Detection dominates the literature, whereas forecasting has fewer studies because public databases are scarce.
- Scope: The review’s main focus is deep-learning research on COVID-19 detection, lung segmentation, and forecasting spread.
- Overall findings: Most reviewed investigations address COVID-19 detection, while forecasting has the fewest studies because available public databases are limited.
- Organization: The tables summarize studies by dataset, imaging modality, sample counts, preprocessing, framework, model, and reported evaluation values.
- Detection studies: The reviewed detection studies use diverse X-ray and CT datasets, including COVID-19, non-COVID-19, pneumonia, and healthy cases.
- Detection studies: Reported detection results include Acc=99.26 for one study and Acc=93 for another study using different datasets and preprocessing settings.
- Architectures: The surveyed methods include CNN, DenseNet, ResNet, VGG, MobileNet, SqueezeNet, 3D-CNN, feature-fusion, GAN, and recurrent approaches.
VI. CHALLENGES
The review identifies limited data, incomplete metadata, uncertain disease dynamics, and constrained hardware as major challenges for DL-based COVID-19 diagnosis, segmentation, and forecasting.
- Data availability: Limited public X-ray and CT databases constrain DL diagnosis and forecasting, leading researchers to use pre-trained networks and leaving forecasting studies relatively scarce.The review also notes that pre-trained models often originate from ImageNet, which differs substantially from medical images.
- Data availability: Missing age and gender information limits the phenotypic data available to DL algorithms, although incorporating it can improve performance.
- Segmentation data: Segmentation datasets often lack manual ground truths, requiring time-consuming radiologist annotation and making model comparison harder.Dedicated segmented-image databases would support stronger model performance and easier comparisons across studies.
- Forecasting: Forecasting prevalence is difficult because COVID-19 remains relatively unknown, may mutate, and depends on demographic, policy, climatic, and social-contact factors.
- Hardware and deployment: Limited access to strong hardware makes practical COVID-19 CADS systems difficult to implement, and most studies do not provide deployable web or Windows software.
VII. CONCLUSION AND FUTURE WORKS
The review emphasizes the importance of fast, accurate COVID-19 diagnosis and surveys deep-learning approaches for diagnosis, segmentation, and forecasting. It identifies limited public datasets as a critical challenge and proposes advanced fusion, zero-shot, attention, and transformer methods for future work.
- COVID-19 can severely damage lung cells, making early and accurate diagnosis important; specialists combine imaging with RT-PCR under WHO guidance.
- The review presents public databases and compares state-of-the-art deep-learning techniques for COVID-19 diagnosis, segmentation, and spread forecasting.
- It surveys advanced attention, transformer, fusion, and graph methods alongside COVID-19 classification, segmentation, and prediction studies.
- The paper introduces COVID-19 diagnosis models, deep-learning toolboxes, and classification algorithms as part of its reviewed methods.
- Limited extensive public databases are identified as a critical challenge because broader datasets could support more accurate and robust detection and prediction models.
- Future research could apply deep feature fusion with medical imaging, zero-shot learning, and diverse attention and transformer models to COVID-19 diagnosis.