| dc.description.abstract | Currently, deep learning is widely used in the field of medicine, which includes radiology. This paper addresses the classification of X-ray images, particularly focusing on the challenge of insufficient images for specific classes, namely COVID-19 and Normal X-ray scans. To tackle this issue, we propose few-shot learning based on different Residual Convolutional Neural Network (CNN) models with varying complexities. This method is tailored for datasets with a small number of samples for a particular class and a larger number for another class, thereby addressing the dataset imbalance issue.
The Residual CNN models utilized in this study are ResNet-50, ResNet-101, and ResNet-152. These architectures were employed to extract features from the images for subsequent classification. The complexity of the models varies, with ResNet-152 being the most complex and ResNet-50 the least.
The results demonstrate promising accuracy, with the highest achieved accuracy of 97.7% for 10 shots of COVID-19 positive X-ray images using the ResNet-101 model. The ResNet-152 model achieved a maximum accuracy of 95.6%. However, on average, the ResNet-101 model outperformed the others. Although the ResNet-50 model provided less accurate results, its lower complexity facilitates faster performance.
Moreover, the study includes t-distributed stochastic neighbor embedding visualization to provide transparency to the proposed solution. This visualization indicates that the system can effectively separate the two classes into distinct clusters.
Overall, the results suggest the efficiency of the proposed solution for addressing the challenge of classifying X-ray images with imbalanced datasets. | en_US |
