Integrated deep learning-based IRACE and convolutional neural networks for chest X-ray image classification

When pre-trained models are applied directly to chest X-ray (CXR) images without appropriate adaptation, they frequently show problems like overfitting, limited generalization, or decreased SE to clinically relevant features because of the unique characteristics of medical data, such as class imbala...

Full description

Saved in:

Bibliographic Details
Published in:	Knowledge-based systems Vol. 329; p. 114293
Main Authors:	Abdel Samee, Nagwan, Houssein, Essam H., Saber, Eman, Hu, Gang, Wang, Mingjing
Format:	Journal Article
Language:	English
Published:	Elsevier B.V 04.11.2025
Subjects:	Chest X-rays (CXR) Convolutional neural networks (CNNs) COVID-19 Deep learning (DL) Iterated race for automatic algorithm configuration (IRACE) Pneumonia COVID-19 Iterated race for automatic algorithm configuration (IRACE) Pneumonia Convolutional neural networks (CNNs) Chest X-rays (CXR) Deep learning (DL)
ISSN:	0950-7051
Online Access:	Get full text
Tags:	Add Tag No Tags, Be the first to tag this record!

Description
Summary:	When pre-trained models are applied directly to chest X-ray (CXR) images without appropriate adaptation, they frequently show problems like overfitting, limited generalization, or decreased SE to clinically relevant features because of the unique characteristics of medical data, such as class imbalance and domain-specific noise. Due to the discrepancy between natural image features (used during pre-training) and radiological image characteristics, studies have shown that such models may perform well on training data but poorly on unseen clinical samples. This study comprehensively evaluates the performance of the fine-tuning method using the Iterated Race for Automatic Algorithm Configuration (IRACE) technique on pre-trained models for several medical imaging CXRs. We select five well-known CNN architectures: MobileNet-v2, EfficientNet-b0, ResNet-50, DenseNet-121, and VGG-19, utilizing the IRACE technique for HPT classification of three CXR datasets. The experimental results indicate that the IRACE technique was generally effective across CXR images, producing noticeable improvements on all models. DenseNet-121 outperformed the other architectures across all metrics, achieving accuracies of 99.83 %, 99.98 %, and 99.87 % on the three CXR datasets, respectively. Additionally, we explored the model detection mechanism by interpreting the classification of radiological images using the Gradient-weighted Class Activation Mapping (Grad-CAM) with Layer-wise Relevance Propagation (LRP) approach for CXR imaging. The results obtained have provided information on how the model classifies CXR images, which can assist radiologists in identifying and evaluating visual characteristics.
ISSN:	0950-7051
DOI:	10.1016/j.knosys.2025.114293