Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{205995,
        author = {Ramratan Sharma and Dileep Kumar Agarwal},
        title = {An Explainable Pneumonia Detection System for Clinical Use on Consumer-Grade Hardware},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {1},
        pages = {8924-8931},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=205995},
        abstract = {Deep learning systems for pneumonia detection from chest radiographs have reached accuracies above 95% on standard benchmarks; however, no published study provides a working application through which a clinician can use the underlying model. This gap persists across more than 15 reviewed publications between 2018 and 2026, none of which discloses inference latency, memory footprint, or a deployable interface. This study presents the design, implementation, and evaluation of a complete clinical decision-support system built around a four-model weighted CNN ensemble (VGG16, EfficientNetB0, DenseNet121, and ResNet50V2), addressing this deployment gap directly. The system is organised into six modules — configuration, preprocessing, model definition, training, evaluation, and a Streamlit-based application layer — and runs entirely offline on a single 4 GB consumer GPU (NVIDIA RTX 3050 Ti). Three design decisions distinguish it from prior work: an adjustable decision threshold exposed directly to the user rather than fixed at the conventional 0.5 cutoff; four-model Grad-CAM spatial explanations rendered alongside the original radiograph for every prediction; and an automated PDF report generator that records patient metadata, per-model probability breakdown, ensemble weighting, and the visual explanation as a single auditable artefact. On the standard 624-image test set, the system achieves 95.19% accuracy, 0.9863 AUC-ROC, and 96.92% recall, with end-to-end inference, explanation generation, and report assembly completing in approximately three seconds per image on the stated hardware. Two representative case studies are presented in full: a true-positive pneumonia case (ensemble probability 92.3%) and a true-negative normal case (ensemble probability 15.1%), illustrating both the per-model disagreement the ensemble resolves and the spatial regions the network attends to. The complete system, including source code organisation and report templates, is documented to a level sufficient for independent reproduction — a standard not met by any directly comparable prior publication.},
        keywords = {Clinical decision support; Pneumonia detection; Explainable AI; Grad-CAM; Streamlit; Software architecture; Consumer GPU; Deployment; Medical imaging},
        month = {June},
        }