MV3Lung-NS: A Neutrosophic-Deep Learning Hybrid Framework for Computer-Aided  Diagnosis of Chest X-Ray Scans

Abdelrhman Fathy Abdelrhman; Ahmed R. Abas

Authors

Abdelrhman Fathy Abdelrhman Department of Computer Science, Faculty of Computer and Informatics, Zagazig University, Zagazig 44519, Egypt
Ahmed R. Abas Department of Computer Science, Faculty of Computer and Informatics, Zagazig University, Zagazig 44519, Egypt

Keywords:

Lung-Infection Diagnosis; Deep Learning; Neutrosophic Sets; Data Augmentation; X-ray Scans, Explainable Artificial Intelligence.

Abstract

The growing adoption of deep learning (DL) for chest X-ray (CXR) diagnosis faces three significant
barriers that this study addresses. First, inherent ambiguities in CXRs - particularly overlapping tissue
intensities between pathological and healthy regions, along with common noise artifacts - create
indeterminate zones where conventional DL models frequently err. Second, the scarcity of high-quality
annotated datasets and persistent class imbalance problems lead to biased and overfitted models. Third,
the opaque decision-making process of DL systems undermines clinical trust, especially in borderline cases.
To resolve these challenges, we implement Neutrosophic Sets (NS) to explicitly quantify and manage
uncertainty at the pixel level through truth-falsity-indeterminacy memberships, particularly effective in
clarifying ambiguous infection boundaries. Simultaneously, we employ radiologist-validated Data
Augmentation (DA) techniques to mitigate data scarcity and imbalance issues. Our results demonstrate NS
filtering enhances model reliability, improving EfficientNetB0 accuracy by 3.13% (94.79% to 97.92%) in
uncertain regions, while DA boosts MobileNetV3Large's generalization capability with a 5.78% accuracy
gain (93.75% to 99.53%). Building on these findings, we propose MV3Lung-NS, an integrated framework
combining NS preprocessing, DA, and MobileNetV3Large that achieves state-of-the-art performance
(99.53% accuracy, 99.65% precision) on pulmonary infection diagnosis. To bridge the interpretability gap,
we implement Explainable AI (XAI) methods including SHapley Additive exPlanations (SHAP), Local
Interpretable Model-agnostic Explanations (LIME), and Gradient-weighted Class Activation Mapping
(Grad-CAM), providing visual evidence that model decisions align with radiological markers of infection.
This work makes dual contributions: advancing neutrosophic theory through empirical validation in
medical imaging and delivering a clinically viable solution that addresses both technical and trust-related
barriers in AI-assisted diagnosis.