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{198365,
        author = {K Nitish V Venkatesh and Mrs. T. Kranthi and Marrapu Tarun and Gorli Praveena and Karanam Swapna Preethi and Shaik Mohammed Marshuk},
        title = {Multi-Stage Deep Learning with Explainability for Chest X-Ray Pathology Detection using DenseNet-121},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {12},
        number = {11},
        pages = {11141-11147},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=198365},
        abstract = {Chest radiography remains one of the most widely deployed and economically accessible modalities in clinical di- agnostics, yet its interpretation is inherently limited by radiologist availability, cumulative fatigue, and measurable inter- observer variability—particularly under high patient-load conditions. This paper presents an intelligent Chest X-Ray Di- agnostic Assistant that addresses these systemic constraints through a cascaded, multi-stage inference architecture. Prior to pathology analysis, each uploaded image traverses a three- level binary validation cascade: MobileNetV2 filters non-medical content, a fine-tuned ResNet18 confirms thoracic anatomy, and a transfer-learned ResNet18 performs a preliminary normal- versus-abnormal triage, collectively ensuring that computation- ally intensive classification is invoked only on radiographically valid inputs. Fourteen thoracic conditions are subsequently identified by DenseNet-121 trained on the NIH ChestX-ray14 benchmark, yielding an average area under the receiver operating characteristic curve (AUROC) of 0.8779. Prediction interpretability is provided through Grad-CAM++ attention heatmaps that spatially localise the image regions driving each classification decision. A Google Gemini large language model (LLM) integration then converts the structured numerical outputs into clinician-readable reports partitioned into summary, findings, and recommendations. The system additionally incorporates a Geoapify-powered specialist-finder that maps detected pathologies to nearby relevant clinicians. Fault-tolerant design across four operational tiers ensures graceful degradation rather than system failure under adverse conditions, establishing the pro- posed assistant as a practically deployable decision-support tool for resource-constrained healthcare environments.},
        keywords = {X-Ray Analysis, DenseNet-121, Convolutional Neural Networks, Binary Classification Pipeline, ResNet18, Transfer Learning, MobileNetV2, Grad-CAM++, Explainable AI, Large Language Model, FastAPI, Thoracic Pathology Detection, NIH ChestX-ray14, Medical Image Analysis},
        month = {April},
        }