Source-linked AI summary
Diagnosing COVID-19 Pneumonia from X-Ray and CT Images using Deep Learning and Transfer Learning Algorithms
Halgurd S. Maghdid, Aras T. Asaad, Kayhan Zrar Ghafoor, Ali Safaa Sadiq, Muhammad Khurram Khan
TL;DR
The study addresses demand for faster COVID-19 diagnosis amid limited publicly available X-ray and CT datasets. It builds a multi-source dataset and applies modified CNN and pretrained AlexNet approaches, with AlexNet achieving around 98% accuracy on COVID-19 X-rays.
Problem
COVID-19 created overwhelming diagnostic demand, while publicly available X-ray and CT image datasets were limited for developing automated detection tools.
Method
The study constructs preprocessed X-ray and CT datasets from multiple sources and applies a modified CNN alongside transfer learning with a modified pretrained AlexNet.
Results
Around 98% overall accuracy was achieved by improved AlexNet on COVID-19 and normal X-ray images, while its performance was weaker when tested on COVID-19 CT images.
Takeaways & Limitations
The publicly available dataset and simple CNN model are intended to support radiologists, data scientists, and research on early COVID-19 diagnosis.
Takeaways & Limitations
The authors caution that the solution is not production-ready because the number of available COVID-19 images is limited.
Abstract
from arXiv · showhide
COVID-19 (also known as 2019 Novel Coronavirus) first emerged in Wuhan, China and spread across the globe with unprecedented effect and has now become the greatest crisis of the modern era. The COVID-19 has proved much more pervasive demands for diagnosis that has driven researchers to develop more intelligent, highly responsive and efficient detection methods. In this work, we focus on proposing AI tools that can be used by radiologists or healthcare professionals to diagnose COVID-19 cases in a quick and accurate manner. However, the lack of a publicly available dataset of X-ray and CT images makes the design of such AI tools a challenging task. To this end, this study aims to build a comprehensive dataset of X-rays and CT scan images from multiple sources as well as provides a simple but an effective COVID-19 detection technique using deep learning and transfer learning algorithms. In this vein, a simple convolution neural network (CNN) and modified pre-trained AlexNet model are applied on the prepared X-rays and CT scan images dataset. The result of the experiments shows that the utilized models can provide accuracy up to 98 % via pre-trained network and 94.1 % accuracy by using the modified CNN.
I. INTRODUCTION
The rapid spread of COVID-19 created overwhelming diagnostic demand, while manual interpretation of CT and X-ray images remained time-consuming and constrained by limited radiologist capacity. The study addresses this need by building a multi-source dataset and applying modified CNN and pre-trained AlexNet models for automated detection.
- Limited radiologist capacity relative to patient numbers can contribute to delayed detection, quarantine, and treatment.
- COVID-19’s rapid spread and overwhelming diagnostic demand prompted the development of faster, more responsive detection methods.
- CT and X-ray interpretation can be time-consuming, especially when infected regions must be manually quantified, creating a need for automatic detection.
- The proposed approach applies a modified CNN and a modified pre-trained deep learning model to COVID-19 diagnosis.
- The study builds a ready-to-use dataset of CT and X-ray images collected from multiple sources.
- The paper presents its dataset and models as contributions intended to support quicker and more accurate COVID-19 diagnosis.
II. BACKGROUND
Prior work used CT-based deep learning systems to support diagnosis, segmentation, and infection quantification, including radiologist-in-the-loop workflows. Reported systems improved diagnostic speed or efficiency, while this paper positions its approach against earlier CNN-based work.
- Earlier research developed deep learning systems using high-resolution CT images to diagnose COVID-19 pneumonia and support radiologists.
- One prior system collected 46,096 anonymous CT images from 106 patients, including 51 laboratory-confirmed COVID-19 cases and 55 controls.
- A deep learning model was reported to achieve performance equivalent to expert radiologists while improving radiologist efficiency.
- The VB-Net system automatically segmented and quantified COVID-19 infection regions and extracted the lung from chest CT scans.
- A human-in-the-loop workflow trained with 249 cases and validated with 300 new cases, reducing annotation time to 4 minutes after three model-update iterations.
- Another study reported 90–92% average accuracy, 100% sensitivity, and 80% specificity using only 50 X-ray images, with limited data noted.
III. THE PROPOSED APPROACH
The proposed approach combines a simple CNN with a modified pre-trained AlexNet to diagnose COVID-19 from pre-processed X-ray and CT images. The models use convolutional feature extraction and classification layers, while AlexNet transfers learned parameters to the new datasets.
- The study uses two approaches: a simple CNN architecture and a transfer-learning algorithm based on AlexNet.
- The modified CNN contains one convolutional layer with 16 filters, batch normalization, ReLU, two fully connected layers, SoftMax, and a classification layer.
- Input images are cropped and resized to remove writing and normalize dimensions across X-ray and CT sources.The CNN input size is described as 224-by-224-by-3, while AlexNet inputs are unified to 227-by-227 with three color channels.
- The convolutional layer extracts image features with sixteen 5 * 5 filters while preserving spatial relationships between pixels.
- Batch normalization stabilizes training, ReLU replaces negative feature values with zero, and fully connected layers classify the extracted features.
- The output labels class 1 as COVID-19 positive and class 0 as a non-COVID-19 normal chest image.
- The modified AlexNet transfers learned weights, biases, and features before training separately on CT scan and X-ray datasets.
A. Datasets
The study assembled COVID-19 X-ray and CT datasets from multiple publicly available sources, then evaluated CNN and pretrained AlexNet models using separate testing sets. The collected data were limited in scale and unevenly distributed across patients and modalities.
- Dataset construction: The dataset combined images from 5 sources, including 170 COVID-19 X-ray images and 361 COVID-19 CT images.The authors selected multiple countries and openly available sources to increase diversity and accessibility.
- Dataset construction: The X-ray images represented 45 patients, while CT images came from 6 BSTI patients and 16 GitHub patients.Some GitHub images were collected from SIRM; the datasets were updated through 18 March 2020.
- Testing setup: Testing used 50 X-ray images and 17 CT images, combining COVID-19 and normal cases from the listed sources.The X-ray test set contained 25 COVID-19 and 25 normal images; the CT test set contained 11 COVID-19 and 6 normal images.
- Model evaluation: The modified CNN achieved 100% sensitivity on X-rays and 90% sensitivity on CT images for suspected COVID-19 cases.These sensitivity values were reported for the study’s two imaging modalities.
- Model evaluation: Pretrained AlexNet correctly identified all COVID-19 X-ray images and 96% of normal X-ray images, but classified only 72% of COVID-19 CT images correctly.The results indicate stronger reported performance on X-rays than on CT images for this model.
- Model comparison: Compared with the CNN in, the proposed CNN matched COVID-19 X-ray sensitivity and performed better on normal X-ray images.The paper also describes the proposed architecture as a simple one-convolutional-layer model with 16 filters sized 5-by-5.
B. Experiments & Results
Experiments evaluated the proposed CNN and modified pretrained AlexNet on X-ray and CT datasets. The modified AlexNet reached 100% validation accuracy and about 98% overall accuracy on the reported X-ray evaluation.
- Reported results: The results section reports performance of the proposed CNN architecture after the experimental setup described previously.The CNN results are summarized in the paper’s performance table.
- AlexNet results: 100% validation accuracy was reached by the modified pretrained AlexNet at 20 epochs and 120 iterations.The paper attributes this to replacing the final layers and using transfer learning rather than training the network from scratch.
- AlexNet results: About 98% overall accuracy was obtained because improved AlexNet identified all COVID-19 X-ray images and 23 of 25 normal X-ray images correctly.The paper reports that the network extracted features from COVID-19 X-ray images and detected lesions and opacities.
V. CONCLUSION
The study introduces CNN and pre-trained AlexNet approaches for COVID-19 detection from publicly available chest X-ray and CT images. The authors report high detection performance while emphasizing that limited image availability prevents production-ready deployment.
- The study introduces a simple CNN and modified pre-trained AlexNet for detecting COVID-19 from chest X-ray and CT images.
- The authors state that high accuracy, sensitivity, and specificity do not establish a production-ready solution because only a limited number of COVID-19 images were available.
- All COVID-19 X-ray images were identified correctly, along with 23 of 25 normal X-ray images.
- The authors propose expanding the X-ray and CT image collection and designing a deeper CNN if needed.