Source-linked AI summary
Covid-19: Automatic detection from X-Ray images utilizing Transfer Learning with Convolutional Neural Networks
Ioannis D. Apostolopoulos, Tzani Bessiana
TL;DR
The paper examines automatic Covid-19 detection from X-Ray images, motivated by diagnostic urgency and cost concerns. It evaluates pre-trained CNNs using transfer learning on publicly sourced images, reporting effective automatic detection while identifying data and diagnostic-scope limitations.
Problem
Covid-19 requires timely diagnosis, while diagnostic costs concern states and patients; the study investigates automatic detection from thoracic X-Rays.
Method
The study evaluates state-of-the-art pre-trained CNNs with transfer learning for Covid-19 diagnosis from thoracic X-Rays.
Results
The results demonstrate the effectiveness of transfer learning with CNNs for automatic detection and essential-feature extraction from Covid-19-related X-Ray images.
Takeaways & Limitations
The work contributes to the possibility of low-cost, rapid, automatic diagnosis that can be clinically exploited in the near future.
Takeaways & Limitations
Further analysis requires more patient data, broader viral and pneumonia cases, and a more holistic patient assessment beyond medical images.
Abstract
from arXiv · showhide
In this study, a dataset of X-Ray images from patients with common pneumonia, Covid-19, and normal incidents was utilized for the automatic detection of the Coronavirus. The aim of the study is to evaluate the performance of state-of-the-art Convolutional Neural Network architectures proposed over recent years for medical image classification. Specifically, the procedure called transfer learning was adopted. With transfer learning, the detection of various abnormalities in small medical image datasets is an achievable target, often yielding remarkable results. The dataset utilized in this experiment is a collection of 1427 X-Ray images. 224 images with confirmed Covid-19, 700 images with confirmed common pneumonia, and 504 images of normal conditions are included. The data was collected from the available X-Ray images on public medical repositories. With transfer learning, an overall accuracy of 97.82% in the detection of Covid-19 is achieved.
1 Introduction
The study addresses the need for timely, affordable Covid-19 diagnosis by evaluating transfer learning with pre-trained CNNs on thoracic X-Rays. Using a small Covid-19 image sample, the approach achieved auspicious results for automatic detection.
- Early automatic Covid-19 diagnosis may support timely quarantine referral, rapid intubation of serious cases, and disease-spread monitoring.
- Diagnostic-test costs can burden states and patients, particularly in private or restricted-access health systems.
- Publicly available X-Rays provide data for studying image patterns that may support automatic Covid-19 diagnosis.
- Deep learning uses CNNs for automatic feature extraction and classification from images, including medical-image classification.
- The study evaluates state-of-the-art pre-trained CNNs for Covid-19 diagnosis from thoracic X-Rays, using transfer learning because only 224 Covid-19 images were available.Transfer learning retains knowledge extracted from one task for an alternative task.
- The reported results demonstrate the effectiveness of transfer learning with CNNs for automatically detecting abnormal X-Ray images related to Covid-19.
2 Methods
The methods combine publicly sourced X-Ray data, transfer learning, CNN configuration, and classification metrics to distinguish Covid-19, common pneumonia, and normal cases. The study also considers data augmentation and cautions that heavy augmentation may create unrealistic images.
- Dataset: X-Ray data were collected from public repositories, including Cohen’s collection, several radiology websites, and common-pneumonia scans.
- Dataset: The dataset contains 224 Covid-19, 700 common-pneumonia, and 504 normal-condition images.
- Transfer Learning: Transfer learning transfers CNN knowledge from a related task to new data, typically when the new dataset is too small for training from scratch.
- Transfer Learning: Feature-extraction transfer learning keeps the pre-trained architecture and learned weights, whereas the alternative modifies architecture or parameters and trains new parameters.
- Data Augmentation: Data augmentation enhances small training sets, while geometric distortions can increase samples or balance dataset sizes.
- Data Augmentation: Heavy data augmentation should be carefully considered because unrealistic images may confuse the CNN.
- CNN Configuration: The CNN experiments tune layer cutoff and the classifier architecture, with late layers made trainable for additional feature extraction.
- Metrics: The study records TP, FN, TN, and FP, and distinguishes three-class accuracy from Covid-19-focused two-class accuracy.
3 Results
The study evaluated CNNs using 10-fold cross-validation and reported accuracy, sensitivity, and specificity in Table 2. Confusion matrices for VGG19 and MobileNet were then examined, with MobileNet outperforming VGG19 in specificity.
- 10-fold cross-validation was used to train and evaluate each CNN.
- Table 2 reports accuracy, sensitivity, and specificity for the CNNs used in transfer learning.
- Some Table 2 measurements are marked unacceptable for real-life problems because of the stated issue.
- Table 3 presents confusion matrices for the two best models, VGG19 and MobileNet.
- MobileNet outperforms VGG19 in specificity and is identified as the most effective model for Covid-19 detection from X-Ray images.
4 Discussion
The discussion concludes that transfer learning with CNNs supports automatic detection and feature extraction from Covid-19-related X-Ray images. It also identifies data diversity and image-only diagnosis as important limitations, while describing potential low-cost and rapid clinical use.
- Transfer learning with CNNs can significantly affect automatic Covid-19 detection and extraction of essential X-Ray features.
- More patient data, especially from Covid-19 cases, is needed for deeper analysis.
- Future models should distinguish Covid-19 from SARS, other viral cases, broader pneumonia cases, and physiological X-Rays.
- The image-only diagnosis does not use a more holistic patient assessment incorporating other potential risk factors.
- The approach contributes to the possibility of low-cost, rapid, automatic diagnosis with potential clinical exploitation.