Comparative Analysis of Transfer Learning and Few-Shot Learning with CNN Architectures for Chest X-Ray Classification under Data Constraints

Abstract

This study focuses on the early and accurate diagnosis of life-threatening lung diseases such as COVID-19, pneumonia, and lung opacity using deep learning. Since deep learning requires large datasets that are often limited in medical imaging, the work applies transfer learning to overcome this challenge. Six pre-trained CNN models—VGG19, VGG16, ResNet50, MobileNetV2, InceptionV3, and DenseNet201—are used to classify chest X-ray images through feature extraction and fine-tuning techniques. In the evaluation phase, a range of classifiers, including Random Forest, K-Nearest Neighbors, Extra Trees, and Decision Tree, were employed to assess the predictive capabilities of the CNN-derived features. The outcomes reveal
insights into the compatibility of these classifiers with different transfer learning strategies. Furthermore, this study delves into the realm of few shot learning, utilizing a limited subset of 15 images from each class. The efficacy of both transfer learning and few-shot learning in the context of this constrained dataset is
examined, shedding light on the adaptability of these techniques to scenarios with limited training samples. The results showcase the strengths and limitations of each approach, providing valuable insights into the intricate landscape of chest X-ray classification.Results show that for the dataset having a total of 3707 images comprising four different classes, the fine-tuned method has outperformed the feature-extracted
method for all the deep learning models executed, giving a high accuracy of 98.89% for the DenseNet201
model with data augmentation and Extra Tree classifier. For the case where only 15 images have been
taken from each of the four classes, Siamese Networks type few-shot learning has outperformed both a base
model and two types of transfer learning models, yielding the best accuracy of 96.84% for the DenseNet
201 model.This work contributes to the ongoing efforts to develop reliable and efficient diagnostic tools
amidst the evolving challenges posed by the recent COVID-19 pandemic.