Source-linked AI summary
Deep Learning for Screening COVID-19 using Chest X-Ray Images
Sanhita Basu, Sushmita Mitra, Nilanjan Saha
TL;DR
The paper addresses the need for a fast alternative to burdensome COVID-19 screening and the scarcity of COVID-19 images for deep-network training. It proposes DETL with pre-trained CNNs and Grad-CAM on four-class chest X-ray data, reporting 90.13% ± 0.14 5-fold cross-validation accuracy and clinically correlated attention regions.
Problem
COVID-19 screening needs fast automated alternatives because PCR testing is burdensome, while available COVID-19 datasets are inadequate for training deep neural networks.
Method
The paper uses Domain Extension Transfer Learning with pre-trained convolutional networks on four-class chest X-ray data and applies Grad-CAM for visual interpretation.
Results
90.13% ± 0.14 5-fold cross-validation accuracy was achieved by VGGNet on the four-class Data-B task.
Takeaways & Limitations
Grad-CAM attention regions strongly correlated with clinical findings, as validated by experts, while the initial results showed promise for larger and more diverse datasets.
Abstract
from arXiv · showhide
With the ever increasing demand for screening millions of prospective "novel coronavirus" or COVID-19 cases, and due to the emergence of high false negatives in the commonly used PCR tests, the necessity for probing an alternative simple screening mechanism of COVID-19 using radiological images (like chest X-Rays) assumes importance. In this scenario, machine learning (ML) and deep learning (DL) offer fast, automated, effective strategies to detect abnormalities and extract key features of the altered lung parenchyma, which may be related to specific signatures of the COVID-19 virus. However, the available COVID-19 datasets are inadequate to train deep neural networks. Therefore, we propose a new concept called domain extension transfer learning (DETL). We employ DETL, with pre-trained deep convolutional neural network, on a related large chest X-Ray dataset that is tuned for classifying between four classes \textit{viz.} $normal$, $pneumonia$, $other\_disease$, and $Covid-19$. A 5-fold cross validation is performed to estimate the feasibility of using chest X-Rays to diagnose COVID-19. The initial results show promise, with the possibility of replication on bigger and more diverse data sets. The overall accuracy was measured as $90.13\% \pm 0.14$. In order to get an idea about the COVID-19 detection transparency, we employed the concept of Gradient Class Activation Map (Grad-CAM) for detecting the regions where the model paid more attention during the classification. This was found to strongly correlate with clinical findings, as validated by experts.
I. INTRODUCTION
COVID-19 screening requires fast, automated approaches because PCR testing is burdensome and COVID-19 datasets are too limited for training deep networks. The paper proposes DETL with chest X-rays and Grad-CAM for interpretable classification.
- PCR screening is described as time-consuming, laborious, complicated, uncomfortable, invasive, and constrained by kit shortages, whereas X-ray machines are widely available and relatively low cost.
- The paper motivates fast automated detection using radiomics and deep learning to process the growing volume of patient data.
- Available COVID-19 datasets are inadequate for training deep neural networks, motivating transfer-learning-based screening from chest X-rays.
- Domain Extension Transfer Learning (DETL) uses a pre-trained convolutional neural network tuned to classify normal, other disease, pneumonia, and COVID-19 X-ray images.
- Grad-CAM is used to identify image regions receiving attention during classification, supporting visually interpretative COVID-19 decisions.
II. THORACIC IMAGING
Thoracic imaging can support COVID-19 diagnosis and monitoring, especially when portable X-rays are needed for critically ill patients. Imaging-based learning targets abnormalities and textural features, but broader datasets remain necessary.
- Portable chest X-rays are useful for critically ill patients who are immobile, while CT is described as more reliable for assessing lung changes.
- Bacterial and viral pneumonia must initially be distinguished from COVID-induced pneumonia, with image normalization across machines considered important.
- COVID-19 imaging findings include ground-glass opacities and consolidation, which may become more frequent and bilateral with longer infection.
- Machine learning can detect abnormalities and extract textural features of altered lung parenchyma associated with COVID-19 signatures.
- A preliminary network used 5,941 chest radiography images from 2,839 cases, but more data was needed to consolidate the global database.
III. MATERIAL AND METHODS
The paper states that it summarizes the datasets and methodology used.
- The section introduces a summary of the datasets used and the methodology employed.
- The stated scope covers both data resources and methodological procedures.
- This section serves as a transition into the paper’s dataset and methodology description.
A. Datasets used
The study combines COVID-19 images from four open-source sources with NIH chest X-ray data to construct two datasets. Data-B supports four-way classification among normal, other disease, pneumonia, and COVID-19.
- 305 COVID-19 X-ray images were acquired from four open-source databases, including collections from Italy, Radiopaedia, J. Paul Cohen et al., and a Spanish hospital.
- The NIH Chest X-ray Dataset contributes 108,948 frontal-view images from 32,717 unique patients covering normal samples and 14 lung, heart, and chest-related diseases.
- Data-A contains two classes, normal and disease, with disease comprising 13 lung, heart, and chest-related diseases excluding pneumonia.
- Data-B contains four classes: normal, other disease, pneumonia, and COVID-19.
- Figure 1 presents sample X-ray images of COVID-19 and pneumonia.
B. Methods
The paper uses domain extension transfer learning to address scarce COVID-19 data by adapting pretrained CNN knowledge to a four-class chest X-ray task.
- Transfer learning: Transfer learning addresses scarce medical data and the computational difficulty of training deep CNNs from scratch.The approach leverages knowledge learned from a source domain and task for a target domain and task with limited annotated data.
- Domain extension transfer learning: Domain extension transfer learning is proposed because natural-image and medical-image domains may contain substantially different high-level features.The method extends the source domain so the target domain contains the source domain, denoted DT ⊃ DS.
- Model adaptation: The study trains a CNN on Data-A for diseased-versus-normal classification before adapting it to four Data-B classes.The four classes are normal, other disease, pneumonia, and Covid-19.
- Model adaptation: The adapted model replaces the pretrained network’s final fully connected layer while retaining the remaining pretrained weights.The replacement layer has one neuron for each of the four target classes.
- Optimization and architectures: The training procedure freezes layers through the last convolutional block, then retrains that block and the fully connected layers using slow-learning-rate SGD.AlexNet, VGGNet, and ResNet are evaluated, with 8, 16, and 50 layers respectively.
IV. EXPERIMENTAL SETUP AND RESULTS
The experiments use CNN training with validation-based model selection and evaluate COVID-19 screening feasibility through five-fold cross-validation on Data-B.
- Training setup: A balanced 20,000-sample validation set was randomly sampled from training data for validation after each epoch and detection of overfitting.The model was trained for 100 epochs, retaining the weights with the best validation accuracy.
- Training setup: The transfer-learning stage uses SGD with a learning rate of 10^-4.
- Results: 90.13% ± 0.14 was the five-fold cross-validation accuracy for VGGNet, compared with 82.98% ± 0.02 for AlexNet and 85.98% ± 0.07 for ResNet.VGGNet was the best-performing network.
- Results: VGGNet correctly classified 99% of Covid and 100% of normal cases in most validation folds.Misclassification occurred between pneumonia and other disease classes, which often co-occur.
V. DECISION VISUALIZATION
Grad-CAM visualizes which feature-map regions influence classification by weighting maps according to class-specific gradients and retaining positive contributions.
- Grad-CAM computation: Grad-CAM uses feature maps from the CNN’s last convolutional layer to produce a discriminative image region.The final map is obtained by weighted summation of the feature maps.
- Grad-CAM computation: The weights quantify each feature map’s importance for the class of interest using gradients of the class score with respect to the feature maps.
- Grad-CAM computation: Grad-CAM applies ReLU to the activation map so only features with positive influence on the class of interest remain.
- Visualization: In the visualization, red regions indicate greatest attention while blue regions indicate least importance.
- Visualization: COVID-19 images highlighted ground-glass opacity, while pneumonia images highlighted typical lung inflammation and normal cases showed no highlighted regions.For other diseases, the model focused on the relevant abnormality.
VI. CONCLUSIONS
The paper presents DETL for alternative COVID-19 screening with chest X-rays and uses Grad-CAM to make model attention visually interpretable.
- Conclusion: The paper presents Domain Extension Transfer Learning for determining characteristic COVID-19 features from chest X-ray images.
- Conclusion: Grad-CAM attention maps strongly correlated with clinical findings, according to expert validation.
- Conclusion: The initial results show promise and may be replicated on bigger and more diverse datasets.
Normal
Figure 4 presents Grad-CAM visual results across different disease classes, including the normal class.
- Normal: Figure 4 reports visual results obtained using Grad-CAM.
- Normal: The figure compares visual results across different disease classes.
- Normal: The normal class is included among the disease classes visualized in Figure 4.