Source-linked AI summary
Detection of Coronavirus (COVID-19) Associated Pneumonia based on Generative Adversarial Networks and a Fine-Tuned Deep Transfer Learning Model using Chest X-ray Dataset
Nour Eldeen M. Khalifa, Mohamed Hamed N. Taha, Aboul Ella Hassanien, Sally Elghamrawy
TL;DR
The paper addresses pneumonia detection from chest X-rays under limited-data conditions, where deep-learning systems require substantial training data and medical annotations are costly. It combines GAN-based image generation with fine-tuned deep transfer learning models, finding that ResNet18 achieved 99% testing accuracy with GAN augmentation. The study uses a dataset of 5,863 X-ray images categorized as normal or pneumonia.
Problem
Deep-learning models need large training datasets, while medical-image annotations are costly and small datasets can impair generalization.
Method
The study combines GAN-based image generation with fine-tuned AlexNet, SqueezeNet, GoogleNet, and ResNet18 transfer-learning models.
Results
99% testing accuracy was achieved by ResNet18 with GAN-based image augmentation.
Takeaways & Limitations
ResNet18 is identified as the most appropriate deep transfer model in the proposed GAN-augmented pneumonia-detection approach.
Takeaways & Limitations
The study assumes that effective AI-model performance depends on sufficient training samples, while using a limited dataset for training.
Abstract
from arXiv · showhide
The COVID-19 coronavirus is one of the devastating viruses according to the world health organization. This novel virus leads to pneumonia, which is an infection that inflames the lungs' air sacs of a human. One of the methods to detect those inflames is by using x-rays for the chest. In this paper, a pneumonia chest x-ray detection based on generative adversarial networks (GAN) with a fine-tuned deep transfer learning for a limited dataset will be presented. The use of GAN positively affects the proposed model robustness and made it immune to the overfitting problem and helps in generating more images from the dataset. The dataset used in this research consists of 5863 X-ray images with two categories: Normal and Pneumonia. This research uses only 10% of the dataset for training data and generates 90% of images using GAN to prove the efficiency of the proposed model. Through the paper, AlexNet, GoogLeNet, Squeeznet, and Resnet18 are selected as deep transfer learning models to detect the pneumonia from chest x-rays. Those models are selected based on their small number of layers on their architectures, which will reflect in reducing the complexity of the models and the consumed memory and time. Using a combination of GAN and deep transfer models proved it is efficiency according to testing accuracy measurement. The research concludes that the Resnet18 is the most appropriate deep transfer model according to testing accuracy measurement and achieved 99% with the other performance metrics such as precision, recall, and F1 score while using GAN as an image augmenter. Finally, a comparison result was carried out at the end of the research with related work which used the same dataset except that this research used only 10% of original dataset. The presented work achieved a superior result than the related work in terms of testing accuracy.
1. Introduction
COVID-19 can cause pneumonia detectable with chest X-rays, motivating AI-assisted diagnosis. However, deep-learning systems require large datasets, while medical-image annotation is costly and small datasets can impair generalization.
- COVID-19 can lead to pneumonia, an infection that inflames the lungs’ air sacs and can be detected using chest X-rays.
- AI and machine-learning techniques are presented as tools to help doctors detect pneumonia accurately and rapidly.
- Deep-learning algorithms require large training datasets, but small medical-image datasets can cause poor generalization and over-synthesis.
- Medical-image annotation is costly and time-consuming because medical experts manually provide ground-truth labels.
- Classic augmentation methods such as rotation modify the geometry or intensity of original images for diagnosis.
2. Related Works
Prior studies used deep-learning approaches on chest X-rays and the same dataset considered here. The paper distinguishes its approach by using GANs to generate additional images and address dataset memorization and overfitting concerns.
- The selected chest-X-ray dataset was chosen because it is available and has been used in previous research for comparison.
- The proposed work differs from related studies by using a generative adversarial network to generate more images and reduce dataset memorization and overfitting.
3. Generative Adversarial Networks and Deep Transfer Learning
GANs jointly train a generator and discriminator through a competitive process, while deep-learning models transform inputs through successive layers. The paper presents these concepts as foundations for image generation and transfer learning.
- Generative Adversarial Networks: GANs train two networks together: a generator that creates images and a discriminator that distinguishes true from false samples.
- GAN architecture: A discriminator output near 0.5 indicates an optimal solution in which true and false samples are not distinguishable.
- GAN architecture: The generator receives random latent input z, while the discriminator outputs the probability that an input image belongs to the true distribution.
- Deep Transfer Learning: Deep-learning models pass data through successive layers, with each layer receiving the previous layer’s output as input.
- Deep Transfer Learning: CNN development produced improved image-classification performance, including reported advances from architectures such as VGG, GoogleNet, and DenseNet.
4. Datasets characteristics
The study uses a pediatric chest-X-ray dataset containing 5,863 JPEG images classified into normal and pneumonia categories.
- The dataset contains 5,863 JPEG chest-X-ray images divided into normal and pneumonia categories.The radiographs came from pediatric patients at Guangzhou Medical Center and underwent low-quality-scan removal.
5. Proposed Model
The proposed model combines GAN-based augmentation with fine-tuned deep transfer learning, using selected compact architectures to classify pneumonia chest X-rays. Experiments compare models with GAN-generated data using testing accuracy and additional performance metrics.
- Model architecture: The model uses GAN augmentation followed by deep transfer learning, with separate preprocessing, training, and testing phases.GAN generates new images during preprocessing; transfer models are then trained and evaluated on the resulting dataset.
- Deep transfer models: AlexNet, SqueezeNet, GoogleNet, and ResNet18 were selected because their relatively few layers reduce training time and computational complexity.The study contrasts these models with larger architectures such as Xception, DenseNet, and InceptionResNet.
- GAN augmentation: The GAN augmentation network uses generator and discriminator components, with convolutional and transposed-convolutional layers forming the augmentation process.The generator includes transposed convolutions, ReLU, batch normalization, and Tanh layers; the discriminator uses convolutional, leaky ReLU, and batch-normalization layers.
- GAN augmentation: Using GAN increased the dataset to 6,240 images and was reported to help overcome overfitting caused by the limited image set.The experiments used 10% of the original dataset while generating the other 90% with GAN.
- Experimental results: 99% testing accuracy was achieved by ResNet18 with GAN, exceeding AlexNet at 96.1%, SqueezeNet at 94.7%, and GoogleNet at 98.6%.ResNet18 also achieved the highest overall two-class testing accuracy and was selected for the proposed model.
- Performance evaluation: ResNet18 achieved the highest reported precision, recall, and F1 score, each at 98.97%, among the evaluated transfer models using GAN.These metrics supported selecting ResNet18 together with GAN for the proposed model.
7 Conclusions and Future Works
The study combines GAN-based image generation with fine-tuned deep transfer learning for pneumonia detection from chest X-rays using a limited dataset. Resnet18 was selected as the most appropriate model, achieving the strongest reported testing accuracy and performance metrics with GAN augmentation.
- Conclusions: The proposed method combines GAN-based image augmentation with fine-tuned deep transfer learning for pneumonia detection from chest X-rays.The evaluated transfer-learning models were AlexNet, GoogLeNet, Squeeznet, and Resnet18.
- Conclusions: 10% of the dataset was used for training, while GAN generated 90% of the images for evaluation of the proposed approach.The dataset contained 5,863 X-ray images in the Normal and Pneumonia categories.
- Conclusions: 99% testing accuracy was achieved by Resnet18 with GAN used as an image augmenter.The study identifies Resnet18 as the most appropriate deep transfer model according to testing accuracy.
- Conclusions: The combined GAN and deep transfer-learning approach achieved higher testing accuracy than related work using the same dataset.The comparison used a smaller portion of the original dataset than the related work described.