Journal of Embedded Systems, Security and Intelligent Systems
Vol 7 No 3 (2026): September 2026

CNN-SVM Hybrid Model with Deep Feature Extraction for Chest X-Ray Image-Based Pneumonia Detection

Indah Purnama Sari (Universitas Muhammadiyah Sumatera Utara)
Yohanni Syahra (Universitas Muhammadiyah Sumatera Utara)
Andi Zulherry (Universitas Muhammadiyah Sumatera Utara)
Mhd. Basri (Universitas Muhammadiyah Sumatera Utara)



Article Info

Publish Date
23 Sep 2026

Abstract

Purpose – This study investigates the effectiveness of a hybrid Convolutional Neural Network–Support Vector Machine (CNN-SVM) framework for pneumonia detection using chest X-ray images, addressing the limitations of conventional softmax-based classification in distinguishing complex radiographic patterns. Design/methods/approach – A publicly available pediatric chest X-ray dataset was used to evaluate the proposed framework. DenseNet121, initialized with pretrained ImageNet weights, was fine-tuned to extract discriminative deep features, which were subsequently classified using an SVM with a Radial Basis Function kernel. Performance was assessed through accuracy, precision, recall, F1-score, receiver operating characteristic analysis, confusion matrices, and cross-validation, with a standalone DenseNet121 classifier serving as the baseline. Findings – The hybrid CNN-SVM framework demonstrated superior classification performance compared with the standalone DenseNet121 model. The integration of deep feature extraction and nonlinear margin-based classification improved the discrimination between pneumonia and normal radiographs. Cross-validation further indicated consistent performance across experimental partitions, although misclassification remained evident in both diagnostic categories. Research implications/limitations – The findings highlight the potential of hybrid deep learning and machine learning approaches to support computer-aided pneumonia screening. However, reliance on a single pediatric dataset, the absence of external clinical validation, and limited model interpretability constrain generalizability and clinical applicability. Originality/value – This study provides an empirical evaluation of DenseNet121-based feature extraction integrated with nonlinear SVM classification, demonstrating its potential advantages over conventional softmax classification while emphasizing comprehensive performance assessment and experimental consistency..

Copyrights © 2026






Journal Info

Abbrev

JESSI

Publisher

Subject

Computer Science & IT

Description

The Journal of Embedded System Security and Intelligent System (JESSI), ISSN/e-ISSN 2745-925X/2722-273X covers all topics of technology in the field of embedded system, computer and network security, and intelligence system as well as innovative and productive ideas related to emerging technology ...