Comparative Analysis of Deep Learning Models for Chest Disease Diagnosis Using Nih X-ray Dataset
Year: 2026
Type: article
Abstract: The integration of deep learning in medical image analysis has significantly advanced computer-aided diagnosis, particularly in chest radiography. However, selecting an optimal convolutional neural network (CNN) architecture for reliable disease classification remains a critical challenge due to data variability, annotation quality, and architectural trade-offs. This study presents a comparative evaluation of three CNN models - DenseNet121, ResNet50, and a custom SimpleCNN - for automated detection of pulmonary infiltrations using a subset of the NIH Chest X-ray dataset. To ensure computational feasibility, only one archive segment was used, and preprocessing included filtering, normalization, and image resizing to 224×224 pixels. Models were trained using cross-entropy loss with the Adam optimizer for five epochs and evaluated on a 20% test split. The performance was assessed using multiple diagnostic metrics essential in medical imaging - accuracy, precision, recall, F1-score, and AUC-ROC - to provide a comprehensive understanding beyond overall accuracy. The ResNet50 model achieved the highest test accuracy and the most balanced trade-off across precision and recall, outperforming DenseNet121 and SimpleCNN. Despite these moderate results, the findings confirm that pre-trained deep architectures generalize more effectively than shallow networks under limited data conditions. The study underscores the impact of dataset size, image resolution, and label quality on diagnostic outcomes. These results form a methodological baseline for further research, where improvements are expected through training on the complete dataset, using full-resolution images, and refining model hyperparameters. Ultimately, this comparative framework contributes to identifying optimal CNN architectures for future clinical diagnostic support systems. Additionally, this study highlights the limitations of small-scale datasets and emphasizes the importance of data augmentation and extended training strategies for improving model performance in medical imaging tasks.
DOI: 10.37943/25decx4995
Authors: Kalizhan Akhmetov, Dinara Kaibassova
Institution: Astana IT University
Language: English
Topic: COVID-19 diagnosis using AI
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: diamond
OpenAlex ID: W7159858211