Radiosight: An EfficientNet-V2 and Grad-CAM Based Chest X-ray Similarity & Explainability System for Pulmonary Disease Diagnosis

Abstract

The requirement of the Chest X-ray reading is necessary in the diagnosis of pulmonary diseases but the lim-itation of the little radiology skills, especially in the rural clinical environment, creates delays and heterogeneity in the diagnosis process. The paper presents n an AI-powered diagnostic support system known as llRadiosight, which combines EfficientNet-V2 to extract features of high resolution Grad-CAM to provide explainability and two similarity engines based on the cosine and Euclidean distance. A model was trained with 1000 images of NIH ChestX-ray (1024 1024 resolution) using augmentation techniques such as flipping and rotation. It hass achieved complete implementation with a web interface developed using Next.js and MongoDB and Firebase storage can be deployed. The suggested system was able to reach a 92.4% accuracy, 0.91 F1-score, and 88% Top-5 similarity precision. Grad-CAM heatmaps increased interpretability and clinician confidence. The potential of radiosight to be adopted in real-life and in particular in resource poor healthcare settings is enormous. These findings are also apparently in line with more articulate strides in elucidable medical imaging systems. [1], [2].

Introduction

The text introduces Radiosight, an AI-powered system designed to improve the diagnosis of chest diseases (such as pneumonia, tuberculosis, COVID-19, pleural effusion, and lung nodules) from chest X-ray images. It addresses the challenge that radiological diagnosis is accurate but limited by workload, shortage of experts, and lack of accessibility in underserved regions. While existing AI systems mainly focus on disease classification, they often lack explainability and clinical usefulness.

Radiosight is proposed as a comprehensive, transparent diagnostic framework that combines multiple capabilities:

• A deep learning backbone using EfficientNet-V2 for high-resolution feature extraction
• Grad-CAM for visual explainability of model decisions
• A similarity-based retrieval system to show past similar cases for clinical reference
• A severity scoring and disease progression module
• A web-based clinical interface for easy deployment and access

The literature review shows that deep learning has significantly improved medical image analysis, but most existing models (such as ResNet, DenseNet, and Vision Transformers) suffer from trade-offs between accuracy, computation cost, and data requirements. EfficientNet-V2 is selected because it offers a strong balance of accuracy, speed, and low computational cost, making it suitable for real-world healthcare settings, especially in resource-limited environments.

The methodology includes preprocessing chest X-rays (resizing, noise reduction, normalization), training on the NIH ChestX-ray dataset, and using EfficientNet-V2 for feature extraction. The system also integrates similarity retrieval using dual metrics (cosine and Euclidean distance), Grad-CAM-based explainability, and a structured pipeline for diagnosis support.

Key innovations of Radiosight include:

• Retrieval of similar past cases for evidence-based diagnosis
• Integration of explainability into the decision-making pipeline
• Severity scoring and disease progression tracking
• Lightweight deployment using modern web technologies
• Optimization for low-resource hardware while maintaining high performance

Experimental comparisons show that EfficientNet-V2 outperforms ResNet, DenseNet, and Vision Transformers in terms of accuracy, training time, and inference speed, achieving about 92.4% accuracy with significantly lower computational cost.

Overall, Radiosight is presented as a unified, clinically usabl

Conclusion

This paper introduces Radiosight, a multi-feature AI-enabled diagnostic support system for the diagnosis of chest X-ray images based on a comprehensive solution to critical challenges in medical imaging AI for the interaction of high-resolution feature extraction, explainable analysis, similarity-based retrieval, and clinical deployment infrastructure. The technical foundation is EfficientNet-V2, where the accuracy of classification is 92.4 In addition to classification, related capacities are combined into a single tool: Grad-CAM visualization offers interpretable heatmaps to allow radiologists to test the logic of the model. Dual-metric similarity retrieval is an evidence-based decision support that presents relevant reference cases. Automated severity scoring and progression tracking extends utility into the realm of treatment tracking generate the kind of outcomes that are vital in the clinical world. There is web based architecture ( Next.js, MongoDB, Fire-base ) posing as an option that allows straight on-premise deployment without a complicated network. This comes first by accessibility-the clinics that have basic connectivity to the internet use Radiosight without specialized software or costly servers. Large scale validation checked robustness: evaluation across datasets (NIH Extended, MIMIC-CXR, PadChest) 88-91 Wider effect is not only the measures of accuracy. Around 4.7 billion patients do not access diagnostic imaging services, and radiologists shortage is critical in Sub-Saharan Africa (0.1 per 100,000, whereas, in North America, it is 12.7). AI diagnostic assistance using inexpensive devices is a direction of the democratization of medical solutions, lessening health inequalities. Time-motion studies showed 40 Nonetheless, it is not without drawbacks: it uses only 1,000 images (not a larger training dataset), is only one-label, has not been conducted in prospective trials, and requires more memory and resources than most cost-restricted community studies. These should be tackled in future by expansion of datasets, multi-label architecture, prospective validation and ultra-lightweight variants. Ethical decision-making, such as algorithmic bias, account-ability, patient privacy, informed consent, etc. need to be addressed. Radiosight provides bias auditing, security frame-works, and explainability mechanisms, but ethical issues change with AI’s saturation in healthcare. In the future, the modular design of Radiosight can be improved: multi-label classification, EHR integration, mobile applications, federated learning, and longitudinal progression modeling. Individual enhancement increases clinical ability and provides new technical and ethical dilemmas necessitating interdisciplinary cooperation. On a final note, Radiosight has proven that it is possible to have useful, explainable, deployable apart medical imaging AI in the implementation of effective architectures, clear reason-ing, and clinical understanding. The system offers template for medical AI serving, not replacing clinicians, augmenting, but not automating expertise, with accessibility as more important than accuracy. With the transition to medical AI, manifold approaches like Radiosight will provide scientists with a clue that devices can make a useful positive change when implemented in technical quality and humanitarian values. It will take time and duration and endeavor to get to prototype phase to extensive impact due to validation, reg-ulatory authority, education, and advancement–but it has the likelihood of saving lives with greater diagnostic access and it is worth it immensely. Radiosight is not the final stop, but a step towards unequivocally accurate, explainable, just, and universally available AI-enhanced healthcare.

References

[1] Akhter, M. A. Rahman, and M. A. Hossain, “Ai-based radiodiagnosis using chest x-rays: A re-view,” Frontiers in Big Data, 2023. [Online]. Available: https://www.frontiersin.org/articles/10.3389/fdata.2023.1120989/full [2] M. Aasem and M. J. Iqbal, “Toward explainable ai in radiology: Ensemble-cam for thoracic disease localiza-tion,” Frontiers in Big Data, 2024. [Online]. Available: https://www.frontiersin.org/articles/10.3389/fdata.2024.1366415/full [3] M. L. Giger, “Computer-aided diagnosis in medical imaging: historical review, current status and future potential,” Computerized Medical Imaging and Graphics, vol. 31, no. 4-5, pp. 198–211, 2007. [4] S. Sun et al., “Multimodal fusion of ecg and chest x-ray with deep learning for improved disease classification,” Computers in Biology and Medicine, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S1687850725007320 [5] I. E. Ihongbe, J. Adebayo, and A. Adekunle, “Evaluating visual explainable ai techniques for chest radiography diagnostics,” PLOS ONE, 2024. [Online]. Available: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0308758 [6] D. L. Pham, C. Xu, and J. L. Prince, “Current methods in medical image segmentation,” Annual Review of Biomedical Engineering, vol. 2, pp. 315–337, 2000. [7] G. H. Dagnaw, T. Tadesse, and G. Alemu, “Explainable artificial intelligence in biomedical imaging: A survey,” arXiv, 2025. [Online]. Available: https://arxiv.org/pdf/2507.07148 [8] N. Veeramani, P. Krishnan, and R. Subramanian, “Nextgen lung disease diagnosis with explainable artificial intelligence,” Scientific Reports, 2025. [Online]. Available: https://www.nature.com/articles/s41598-025-07603-4 [9] P. Kaushik, S. Verma, and D. Gupta, “Radiological feature fusion and explainable ai for pneumonia detection,” Artificial Intelligence in Medicine, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2590123025012976 [10] S. P. Koyyada, A. Sethi, and V. Singh, “An explainable artificial intelligence model for identifying radiological patterns in chest x-rays,” Computers in Biology and Medicine, 2023. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S2772442523000734 [11] A. K. Jain, Fundamentals of Digital Image Processing. Prentice-Hall, 2002. [12] S. Biswas et al., “Flpnexainet: Federated deep learning and explainable ai for pneumonia prediction using chest x-rays,” PLOS ONE, 2025. [Online]. Available: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0324957 [13] Z. H. Zhou et al., “Deep convolutional neural networks for chest pathology detection,” Pattern Recognition, vol. 48, no. 6, pp. 2032–2043, 2015. [14] M. T. Zamir, R. Khan, and S. Ali, “Explainable ai-driven analysis of radiology reports and chest x-rays,” JMIR Formative Research, 2025. [Online]. Available: https://formative.jmir.org/2025/1/e77482 [15] X. Fu et al., “Explainable hybrid transformer for multi-classification of lung diseases from chest x-ray images,” Scientific Reports, 2025. [Online]. Available: https://www.nature.com/articles/s41598-025-90607- [16] D. Diwakar et al., “Interpretable chest x-ray localization using principal explainable ai framework,” Artificial Intelligence in Medicine, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/abs/pii/S0952197625023668 [17] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,” in Proceedings of the IEEE, vol. 86, no. 11, 1998, pp. 2278–2324.

Copyright

Copyright © 2026 Ashwani Kumar Yadav, Divyanshu Tiwari, Anurag Singh, Rishabh Mishra, Dr. Seema Maitrey. This 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.