A hybrid dense convolutional network and fuzzy inference system for pneumonia diagnosis with dynamic symptom tracking.
Pneumonia is a major cause of mortality among children under five and adults over 65, especially in low-resource settings where access to skilled radiologists is limited. Accurate and early diagnosis is essential, but is often hindered by subjective interpretation and variability in its symptoms.
This study aims to develop a hybrid Artificial Intelligence (AI) based pneumonia diagnosis system that integrates Deep Learning (DL) confidence scores, DenseNet201 with Capsule Network (CapsNet), Mamdani-style fuzzy inference, and a dynamic symptom adjustment mechanism to enhance diagnostic accuracy, transparency, and clinical usability.
The system was evaluated using 17,229 labelled chest X-ray images across multiple cross-validation techniques: Stratified, k-fold, Bootstrap, and Monte Carlo methods, each with five dataset iterations or folds. DenseNet was used to extract spatial features, while CapsNet preserved spatial orientation and hierarchical relationships. A DL based confidence score was generated and used as a fuzzy membership input to support classification in borderline cases, where severity scores were nearly tied, and the confidence score guided the final decision. A dynamic adjustment algorithm further refined symptom severity by incorporating recent trends in patient data.
The DenseNet201 + CapsNet architecture achieved the highest performance in the 5th fold of stratified cross-validation, with a test accuracy of 99.01%. The model also demonstrated strong generalization, with a weighted precision, recall and F1-score of 0.9878, 0.9874, and 0.9876, respectively, across all classes. The paired t-test confirmed that the CapsNet-based approach outperformed traditional fully connected layers, and the fuzzy logic system effectively handled ambiguous cases using DL confidence. The dynamic membership mechanism showed strong adaptability for real-time symptom tracking.
This hybrid model offers a robust, interpretable, and clinically relevant decision-support tool for pneumonia diagnosis. It bridges high-performance AI with real-world medical decision-making, especially in settings with limited radiological expertise.
This study aims to develop a hybrid Artificial Intelligence (AI) based pneumonia diagnosis system that integrates Deep Learning (DL) confidence scores, DenseNet201 with Capsule Network (CapsNet), Mamdani-style fuzzy inference, and a dynamic symptom adjustment mechanism to enhance diagnostic accuracy, transparency, and clinical usability.
The system was evaluated using 17,229 labelled chest X-ray images across multiple cross-validation techniques: Stratified, k-fold, Bootstrap, and Monte Carlo methods, each with five dataset iterations or folds. DenseNet was used to extract spatial features, while CapsNet preserved spatial orientation and hierarchical relationships. A DL based confidence score was generated and used as a fuzzy membership input to support classification in borderline cases, where severity scores were nearly tied, and the confidence score guided the final decision. A dynamic adjustment algorithm further refined symptom severity by incorporating recent trends in patient data.
The DenseNet201 + CapsNet architecture achieved the highest performance in the 5th fold of stratified cross-validation, with a test accuracy of 99.01%. The model also demonstrated strong generalization, with a weighted precision, recall and F1-score of 0.9878, 0.9874, and 0.9876, respectively, across all classes. The paired t-test confirmed that the CapsNet-based approach outperformed traditional fully connected layers, and the fuzzy logic system effectively handled ambiguous cases using DL confidence. The dynamic membership mechanism showed strong adaptability for real-time symptom tracking.
This hybrid model offers a robust, interpretable, and clinically relevant decision-support tool for pneumonia diagnosis. It bridges high-performance AI with real-world medical decision-making, especially in settings with limited radiological expertise.