Source-linked AI summary
Coronavirus (COVID-19) Classification using CT Images by Machine Learning Methods
Mucahid Barstugan, Umut Ozkaya, Saban Ozturk
TL;DR
The paper addresses early COVID-19 diagnosis from CT images, where characteristic findings motivate timely detection. It creates patch-based datasets, extracts features, and uses SVM classification. The best reported accuracy is 99.68% with 10-fold cross-validation, but the method should be tested on another coronavirus CT dataset.
Problem
Clinical experts need lung CT images to diagnose COVID-19 in its early phase because infected images show characteristic findings.
Method
The study forms four datasets from 150 CT images, applies feature extraction, and classifies the extracted features with SVM.
Results
99.68% classification accuracy was obtained with 10-fold cross-validation.
Takeaways & Limitations
The findings indicate that the proposed method could be used as an assistant system for diagnosing COVID-19.
Takeaways & Limitations
The proposed method should be tested on another coronavirus CT image dataset.
Abstract
from arXiv · showhide
This study presents early phase detection of Coronavirus (COVID-19), which is named by World Health Organization (WHO), by machine learning methods. The detection process was implemented on abdominal Computed Tomography (CT) images. The expert radiologists detected from CT images that COVID-19 shows different behaviours from other viral pneumonia. Therefore, the clinical experts specify that COVİD-19 virus needs to be diagnosed in early phase. For detection of the COVID-19, four different datasets were formed by taking patches sized as 16x16, 32x32, 48x48, 64x64 from 150 CT images. The feature extraction process was applied to patches to increase the classification performance. Grey Level Co-occurrence Matrix (GLCM), Local Directional Pattern (LDP), Grey Level Run Length Matrix (GLRLM), Grey-Level Size Zone Matrix (GLSZM), and Discrete Wavelet Transform (DWT) algorithms were used as feature extraction methods. Support Vector Machines (SVM) classified the extracted features. 2-fold, 5-fold and 10-fold cross-validations were implemented during the classification process. Sensitivity, specificity, accuracy, precision, and F-score metrics were used to evaluate the classification performance. The best classification accuracy was obtained as 99.68% with 10-fold cross-validation and GLSZM feature extraction method.
1. INTRODUCTION
The study addresses early COVID-19 detection from CT images, motivated by characteristic imaging patterns and the need for timely diagnosis. It uses patch-based feature extraction with SVM classification and reports that the method could assist diagnosis.
- COVID-19 produces characteristic CT-image findings, and clinical experts need lung CT images for early diagnosis.
- Prior work classified COVID-19, Influenza-A viral pneumonia, and healthy cases with a 3D deep-learning model, achieving 87.6% overall accuracy.
- A previous system segmented and quantified infected lung regions, achieving a 91.6% Dice similarity coefficient and reducing delineation time to four minutes.
- This study used 150 CT images and created four datasets before classification.
- Feature extraction methods and SVM were used to classify coronavirus and non-coronavirus images.
- The findings indicate that the proposed method could be used as an assistant system for diagnosing COVID-19.
2. MATERIAL
The material comprises abdominal CT images converted into four patch subsets, with infected regions selected and labelled for classification. Images from different CT tools introduce grey-level variation that disadvantages classification.
- The dataset contains 150 abdominal CT images from 53 infected cases, and patches were cropped from selected regions.
- Four different patch subsets were created from the cropped patch regions.
- Images acquired from different CT tools make classification difficult because grey-level meanings vary between images.
- Visual Features of Dataset: Figure 1 shows labelled infected areas in images acquired from different CT tools, where grey levels differ.
- Visual Features of Dataset: Figure 2 presents patch regions and sample patches from the four subsets for infected and non-infected situations.
3. METHOD
The study uses a two-stage SVM classification process, comparing direct patch classification with classification after texture or frequency-based feature extraction. Multiple patch subsets and cross-validation schemes support evaluation of the extracted representations.
- Stage 1: Stage 1 vectorizes four image subsets without feature extraction and classifies them using SVM.The subsets correspond to the study’s patch-based datasets.
- Stage 2: Stage 2 extracts features from all subsets using GLCM, LDP, GLRLM, GLSZM, or DWT before SVM classification.The five methods include statistical, directional, run-length, size-zone, and wavelet-based representations.
- Evaluation: The experiments use 2-fold, 5-fold, and 10-fold cross-validation, with mean classification results obtained after validation.Figure 3 summarizes the two-stage classification process.
- Feature extraction: GLRLM represents grey-level runs with an L×R matrix and produces a 1x7 texture-feature vector.Its features include short-run emphasis, long-run emphasis, grey-level non-uniformity, and run-length non-uniformity.
- Feature extraction: GLSZM is described as a developed version of GLRLM that produces a 1x13 feature vector for classifier input.Its features include zone emphasis, grey-level non-uniformity, size-zone non-uniformity, zone percentage, and variance measures.
4. EXPERIMENTAL RESULTS
The study compares two-stage SVM classification with and without feature extraction across four patch subsets, using cross-validation and five evaluation metrics. The strongest reported result was 99.68% accuracy for Subset 1 with GLSZM features and 10-fold cross-validation.
- Classification procedure: Stage 1 classified image subsets without feature extraction, while Stage 2 classified features extracted from the subsets.Five feature extraction methods were used in Stage 2.
- Evaluation: Sensitivity, specificity, accuracy, precision, and F-score evaluated classification performance.The metrics were computed from true-positive, true-negative, false-positive, and false-negative counts.
- Subset 1: 99.68% accuracy was the best result for Subset 1 in Stage 2 with 10-fold cross-validation and GLSZM feature extraction.Subset 1 contained 5,912 non-infected and 6,940 infected patches.
- Subset 2: 99.37% accuracy was the best result for Subset 2 in Stage 2 with 10-fold cross-validation and DWT feature extraction.Subset 2 contained 942 non-infected and 1,122 infected patches.
- Subset 3: 99.64% accuracy was the best result for Subset 3 in Stage 2 with 10-fold cross-validation and DWT feature extraction.The passage identifies this as the best result for Subset 3.
- Classifier implementation: The extracted patches were converted into feature vectors and classified using five different SVM structures, with mean performance obtained after classification.The figure passage describes 32x32 patches and GLSZM feature extraction for infected-patch detection.
5. DISCUSSION and CONCLUSION
The discussion frames abdominal CT classification as a response to COVID-19’s distinct CT findings and the need for more machine-learning studies. The manually formed dataset achieved 99.68% classification accuracy, but the method should be tested on another coronavirus CT dataset.
- Motivation: COVID-19 CT images show findings that differ from other viral pneumonia, including bronchiectasis, lesion swelling, and different shadowiness.The study presents these findings as relevant to COVID-19 diagnosis.
- Method rationale: Five feature extraction methods were used because the image set contained images acquired with different CT tools, seeking features that separate infected patches.The methods were applied to the coronavirus image set.
- Conclusion: 99.68% classification accuracy was achieved on the manually formed dataset, while the proposed method should be tested on another coronavirus CT image dataset.The recommendation defines the stated validation boundary.
- Research context: The study calls for more classification and segmentation research using abdominal CT, chest X-ray, and blood-test data as such datasets become available.It also identifies dataset diversification as a need.