Source-linked AI summary
Coronavirus (COVID-19) Classification using Deep Features Fusion and Ranking Technique
Umut Ozkaya, Saban Ozturk, Mucahid Barstugan
TL;DR
COVID-19 diagnosis requires fast and accurate approaches, while CT images can show varying disease characteristics and imaging-device effects. The paper fuses and ranks deep features from CT patches before SVM classification, achieving its strongest reported performance on Subset-2.
Problem
Fast and accurate COVID-19 diagnosis is needed, while CT images can exhibit varying characteristics and device-related grey-level differences.
Method
The method extracts deep features from CT patches with pre-trained CNNs, fuses and ranks them, and classifies the resulting data using an SVM.
Results
98.27% accuracy, 98.93% sensitivity, 97.60% specificity, 97.63% precision, 98.28% F1-score and 96.54% MCC were obtained for Subset-2.
Takeaways & Limitations
The proposed method is reported as usable for early diagnosis of COVID-19 cases.
Takeaways & Limitations
Different CT devices produce different infected-region grey levels, which affects classification and performance.
Abstract
from arXiv · showhide
Coronavirus (COVID-19) emerged towards the end of 2019. World Health Organization (WHO) was identified it as a global epidemic. Consensus occurred in the opinion that using Computerized Tomography (CT) techniques for early diagnosis of pandemic disease gives both fast and accurate results. It was stated by expert radiologists that COVID-19 displays different behaviours in CT images. In this study, a novel method was proposed as fusing and ranking deep features to detect COVID-19 in early phase. 16x16 (Subset-1) and 32x32 (Subset-2) patches were obtained from 150 CT images to generate sub-datasets. Within the scope of the proposed method, 3000 patch images have been labelled as CoVID-19 and No finding for using in training and testing phase. Feature fusion and ranking method have been applied in order to increase the performance of the proposed method. Then, the processed data was classified with a Support Vector Machine (SVM). According to other pre-trained Convolutional Neural Network (CNN) models used in transfer learning, the proposed method shows high performance on Subset-2 with 98.27% accuracy, 98.93% sensitivity, 97.60% specificity, 97.63% precision, 98.28% F1-score and 96.54% Matthews Correlation Coefficient (MCC) metrics.
1. INTRODUCTION
The introduction motivates fast and accurate COVID-19 diagnosis because symptoms and laboratory testing can be insufficient or slow. It presents a patch-based deep-feature fusion and ranking method using SVM classification for CT images.
- Fast and accurate diagnosis is necessary because laboratory tests can take longer and have a high margin of error.
- Radiological imaging is introduced as an approach for early detection of suspicious COVID-19 cases, with reported accuracy of 90%.
- Pre-trained CNNs extract deep features that are fused and ranked before classification with a Support Vector Machine.
- The proposed method is presented as applicable to early diagnosis of COVID-19 cases.
2. MATERIAL
The material comprises CT-derived patches from infected and non-infected regions, generated with different imaging devices and organized into two patch datasets. Device variation produces differing grey levels that affect classification.
- Patch images are obtained from infected and non-infected regions of CT images and characterized in two different patch datasets.
- Different CT tools and devices produce COVID-19 regions with different grey levels in the infected images.
- The study extracts 16 × 16 and 32 × 32 patches with different characteristics from the CT images.
- The patch-generation process is illustrated in Figure 1.
3. METHOD
The method section introduces deep learning and deep neural-network concepts used to obtain image features. It describes multilayer representations and CNN-based processing as the methodological background.
- 3.1. Deep Learning: In convolutional neural networks, classification takes place in fully connected and softmax layers.
- 3.1. Deep Learning: Deep learning has been applied to image, video, speech, sound, text, and object-processing tasks.
- 3.1. Deep Learning: Deep learning uses multilayer machine-learning calculations and artificial neural networks to process data through multiple layers.
- 3.1. Deep Learning: Deep Belief Networks contain multiple hidden-node layers whose learned representations act as feature detectors.
- 3.1. Deep Learning: Deep Belief Networks can be further trained for classification after an unsupervised pre-training phase.
3.2. Convolutional Neural Network
The section introduces CNNs as networks built from convolution, activation, pooling, normalization, fully connected, and softmax layers. Convolution combines an input with a kernel to produce feature maps, while pooling summarizes nearby outputs and reduces representation size.
- CNN architecture: CNNs use convolutional, ReLU, pooling, normalization, fully connected, and softmax layers in their architectures.The section describes convolutional, pooling, and fully connected layers as the main ConvNet layer types.
- Convolution: Convolution combines multidimensional inputs with kernels to produce feature maps through sliding operations.For images, the input and kernel are represented as multidimensional arrays, and the kernel is shifted across the input.
- Convolution properties: Convolution uses sparse interactions and shared parameters, reducing the number of learned parameters and computation.The section also notes that convolution can handle variable-sized inputs and preserve edge information through padding.
- Pooling: Pooling replaces local outputs with summary statistics, including maximum, mean, or minimum values, and reduces matrix size.Figure 3 presents 2 × 2 maximum-pooling and mean-pooling operations.
- Activation: ReLU computes F(x) = max(0, x), but excessively high learning rates can leave 40% of the network dead.The cited passage describes this as a less frequent outcome when the activation function is appropriately adjusted.
3.3. Feature Fusion and Ranking Technique
The method fuses deep features from multiple pretrained CNNs and ranks them to reduce insufficiency, redundancy, computational complexity, and processing time.
- Feature fusion: VGG-16, GoogleNet, and ResNet-50 features are fused into higher-dimensional feature representations.Fusion is intended to reduce the effect of insufficient features from any single CNN network.
- Motivation: Feature ranking addresses correlations and excessive information that increase processing time and computational complexity.The section motivates ranking after noting correlation and redundancy among fused features.
- Feature ranking: A t-test ranks fused features by statistically comparing feature differences and considering feature frequency.The ranking process accounts for repeated or closely related features in the feature vector.
3.4. Support Vector Machines (SVMs)
The processed feature vectors are classified with a binary SVM using a linear kernel and squared hinge-loss objective.
- Classifier: A binary SVM classifier is trained after feature fusion and ranking.The SVM maps features into a space where kernel functions can support classification.
- Optimization: The classifier uses a linear kernel and minimizes squared hinge loss.The loss function is given in Equation 10.
- Loss parameters: The fusion-and-ranking feature vector is denoted x_n, while hyperparameter C determines the wrong-classification penalty.These terms are specified in the explanation of the loss function.
3.5. Proposed Method
The proposed method classifies COVID-19 texture using two patch-size datasets, pretrained CNN feature extraction, feature fusion and ranking, and SVM classification.
- Feature extraction: Pretrained VGG-16, GoogleNet, and ResNet-50 networks extract features from the patch images.The networks are trained separately for Subset-1 and Subset-2 before feature extraction.
- Fusion and ranking: Feature vectors from the CNNs are fused, correlated features are considered, and a t-test ranks the resulting features.Features close to one another are eliminated according to feature frequency during ranking.
- Classification: The ranked fused features are evaluated with an SVM classifier for COVID-19 classification.The method workflow is visualized in Figure 4.
- Dataset design: The study classifies COVID-19 texture using two datasets with different patch sizes.The section explicitly identifies the two-dataset design as part of the proposed method.
- Evaluation: Sensitivity, specificity, accuracy, precision, F-score, and MCC are used to evaluate the method.The supplied metric definitions include true-positive, true-negative, false-positive, and false-negative terms.
4.1. Classification Results of Subset 1
Subset-1 contains balanced 16 × 16 CT-patch classes, with 75% of the data used for training and 25% for testing. Classification results are reported against pre-trained CNN networks.
- Dataset and evaluation setup: Subset-1 contains 6000 balanced 16 × 16 CT patches.The data distribution between COVID-19 and No finding classes is equal.
- Dataset and evaluation setup: 75% of Subset-1 was used for training and 25% for testing.
- Comparative results: Table 2 reports comparative classification results for pre-trained CNN networks and the proposed method.
4.2. Classification Results of Subset 2
Subset-2 consists of balanced 32 × 32 CT patches from COVID-19 and No finding classes, with comparative classification results reported in Table 3.
- Dataset and evaluation setup: Subset-2 includes 3000 COVID-19 and 3000 No finding 32 × 32 CT patches.
- Comparative results: Table 3 presents the comparative classification results for Subset-2.
4.3. Performance Evaluation
The proposed method was evaluated using comparative metrics and confusion matrices on both subsets. It achieved its strongest reported overall performance on Subset-2, while class-level results were also reported for each subset.
- Subset-1 results: 95.60% was the best reported performance for the proposed method in Subset-1.
- Subset-1 results: 98.93% sensitivity was achieved by ResNet-50, while GoogleNet reached 98.93% specificity and 98.75% precision.
- Subset-1 results: 95.50% F1-score and 91.29% MCC made the proposed method the most successful among the pre-trained CNN structures.
- Subset-2 results: 98.27% accuracy was achieved by the proposed method on Subset-2.
- Subset-2 results: 98.28% F1-score and 96.54% MCC were the highest metric performances achieved by the proposed method on Subset-2.
- Confusion-matrix results: In Subset-1, COVID-19 and No finding class accuracies were 97.9% and 93.3%, respectively.
- Confusion-matrix results: In Subset-2, COVID-19 and No finding class accuracies were 97.6% and 98.9%, respectively.
5. DISCUSSION and CONCLUSION
The study addresses rapid COVID-19 detection in CT images using fused and ranked deep features from pretrained CNNs. The proposed method achieved stronger reported performance on Subset-2 than Subset-1, although differing CT devices complicated image analysis.
- Motivation: COVID-19 was difficult to detect quickly, motivating CT-based analysis for more accurate diagnosis.
- Limitation: Different CT devices produced different grey-level characteristics, complicating image analysis.
- Method: Deep features from pretrained CNN networks were fused and ranked after random patches were taken from CT images.
- Results: 95.60% accuracy, 95.60% sensitivity, 93.33% specificity, 97.87% precision, 97.77% F1-score and 91.29% MCC were obtained for Subset-1.
- Results: 98.27% accuracy, 98.93% sensitivity, 97.60% specificity, 97.63% precision, 98.28% F1-score and 96.54% MCC were obtained for Subset-2.