Source-linked AI summary
Improving performance of CNN to predict likelihood of COVID-19 using chest X-ray images with preprocessing algorithms
Morteza Heidari, Seyedehnafiseh Mirniaharikandehei, Abolfazl Zargari Khuzani, Gopichandh Danala, Yuchen Qiu, Bin Zheng
TL;DR
Accurate and efficient chest X-ray detection of COVID-19 pneumonia remains challenging, particularly when distinguishing it from other community-acquired pneumonia. The study develops a preprocessing-enhanced VGG16 transfer-learning CAD scheme for three-class classification, achieving 93.9% overall accuracy and 98.6% COVID-19 detection accuracy. The authors conclude that the approach is feasible as a radiologist decision-support tool, while requiring validation on larger, more diverse datasets.
Problem
Accurately and efficiently detecting COVID-19 pneumonia and distinguishing it from other community-acquired pneumonia in chest X-rays remains challenging.
Method
A VGG16 transfer-learning CAD scheme preprocesses chest X-rays and classifies them as COVID-19 pneumonia, other pneumonia, or normal.
Results
93.9% overall accuracy was achieved for three-class classification, while COVID-19 detection reached 98.6% overall accuracy.
Takeaways & Limitations
The study demonstrates feasibility of using a chest X-ray CAD scheme as a decision-support tool for detecting and diagnosing COVID-19 pneumonia.
Takeaways & Limitations
Performance and robustness require further validation on other large and diverse image databases because the COVID-19 cases are heterogeneous and limited in number.
Abstract
from arXiv · showhide
As the rapid spread of coronavirus disease (COVID-19) worldwide, chest X-ray radiography has also been used to detect COVID-19 infected pneumonia and assess its severity or monitor its prognosis in the hospitals due to its low cost, low radiation dose, and wide accessibility. However, how to more accurately and efficiently detect COVID-19 infected pneumonia and distinguish it from other community-acquired pneumonia remains a challenge. In order to address this challenge, we in this study develop and test a new computer-aided diagnosis (CAD) scheme. It includes several image pre-processing algorithms to remove diaphragms, normalize image contrast-to-noise ratio, and generate three input images, then links to a transfer learning based convolutional neural network (a VGG16 based CNN model) to classify chest X-ray images into three classes of COVID-19 infected pneumonia, other community-acquired pneumonia and normal (non-pneumonia) cases. To this purpose, a publicly available dataset of 8,474 chest X-ray images is used, which includes 415 confirmed COVID-19 infected pneumonia, 5,179 community-acquired pneumonia, and 2,880 non-pneumonia cases. The dataset is divided into two subsets with 90% and 10% of images in each subset to train and test the CNN-based CAD scheme. The testing results achieve 94.0% of overall accuracy in classifying three classes and 98.6% accuracy in detecting Covid-19 infected cases. Thus, the study demonstrates the feasibility of developing a CAD scheme of chest X-ray images and providing radiologists useful decision-making supporting tools in detecting and diagnosis of COVID-19 infected pneumonia.
1. INTRODUCTION
COVID-19 created a need for accurate, efficient chest X-ray analysis that distinguishes infected pneumonia from other pneumonia. The study addresses this with a VGG16 transfer-learning CAD scheme supported by image preprocessing.
- Chest X-ray radiography offers low cost, low radiation dose, ease of operation, and wide accessibility for COVID-19 assessment.
- High-volume interpretation remains difficult because subtle COVID-19 pneumonia must be distinguished from other community-acquired pneumonia.
- CAD systems combine medical-image processing and machine learning to analyze disease-related characteristics and support radiologists’ detection and diagnosis.
- Prior COVID-19 CAD studies used deep-learning feature extraction, SVM classification, transfer learning, and CNN models across medical-image applications.
- The proposed scheme uses a pretrained VGG16 network with transfer learning to classify chest X-rays into COVID-19 pneumonia, other pneumonia, and normal cases.
2. MATERIALS AND METHOD
The study builds a three-class chest X-ray CAD scheme by combining diaphragm removal, image enhancement, transfer learning with VGG16, and class-balanced training/testing procedures.
- Dataset: 8,474 chest X-ray images comprise 415 COVID-19, 5,179 community-acquired pneumonia, and 2,880 normal cases.The dataset uses posteroanterior chest views from publicly available medical repositories.
- Image Pre-Processing: The blob discovery step segments high-intensity pixels and removes the largest connected region to suppress the diaphragm.It thresholds pixels at T = V_min + 0.9 × (V_max − V_min), then applies opening, closing, and dilation filters.
- Image Pre-Processing: Three processed images—diaphragm-removed, bilateral-filtered, and histogram-equalized—populate the VGG16 model’s three RGB channels.Each channel is resized to 224×224 pixels for compatibility with the pretrained model.
- Transfer Learning: The transfer-learning model retains VGG16 blocks 1–5, replaces the upper layers, and adds fully connected layers with ReLU and a three-class Softmax output.VGG16 was pretrained on ImageNet; the resulting model has 21,170,755 trainable parameters.
- Model Training and Testing: The dataset is split into independent training and testing subsets using a 90%/10% ratio within each class.The testing subset contains 42 COVID-19, 518 community-acquired pneumonia, and 288 normal cases.
- Performance Assessment: Performance is assessed for both three-class classification and binary COVID-19 versus non-COVID-19 classification using accuracy and additional confusion-matrix metrics.The three-class evaluation includes macro and weighted accuracy, precision, recall, F1-score, and Cohen’s Kappa; the binary evaluation includes TP, FN, TN, and FP-based measures.
3. RESULTS
The VGG16-based CNN achieved high performance in both three-class classification and binary COVID-19 detection. Preprocessing improved comparative model performance, while the selected model correctly classified all COVID-19 testing cases with limited false positives.
- COVID-19 detection: All COVID-19 cases were classified correctly, with 12 false positives comprising 4 normal cases and 8 community-acquired pneumonia cases.The confusion matrix used the model saved at the highest validation accuracy.
- Three-class classification: 93.9% overall accuracy was achieved on 848 testing cases for three-class classification, with 796 cases correctly classified and Cohen’s kappa of 0.88.The 95% confidence interval for overall accuracy was [0.92,0.96].
- COVID-19 detection: 100% sensitivity (42/42), 98.5% specificity (794/806), and 98.6% overall accuracy (836/848) were achieved for binary COVID-19 detection.Normal and community-acquired pneumonia images were grouped as negative, while COVID-19 cases formed the positive class.
- Preprocessing comparison: 91% accuracy and a Cohen’s kappa score of 0.82 were obtained by the filter base model, compared with 88% and 0.75 for the simple model.These models evaluated the effect of filtering-related preprocessing against simply duplicating the original X-ray image across channels.
4. DISCUSSION
The discussion attributes performance gains to transfer learning, multi-channel preprocessing, diaphragm removal, and data augmentation. The authors caution that validation on larger, more diverse datasets and further preprocessing development remain necessary.
- Transfer learning: Transfer learning was used because the 8,474-image dataset was unbalanced and contained only 415 COVID-19 cases, limiting robust retraining of VGG16’s over 138 million parameters.The original parameters were trained on more than 14 million ImageNet images.
- Multi-channel preprocessing: Using three different input images increased overall classification accuracy 3.2% from 91.0% to 93.9% and Cohen’s kappa 7.3% from 0.82 to 0.88.The inputs comprised the original, bilateral-filtered, and histogram-equalized images assigned to the CNN’s three channels.
- Diaphragm removal: Removing diaphragm regions changed overall accuracy from 87.6% to 93.9% and Cohen’s kappa from 0.74 to 0.88.The authors report this as increases of 7.2% in accuracy and 18.9% in Cohen’s kappa.
- Data augmentation: Without data augmentation, overall classification accuracy fell to around 78%.The authors describe augmentation as essential for increasing the training dataset size.
- Limitations: Performance and robustness require further validation using other large and diverse image databases, and the tested preprocessing methods may not be optimal.The authors also call for improved algorithms to remove the diaphragm and regions outside the lungs.