A Full-Stack, Explainable Chest X-ray Analysis Workspace Using DenseNet121: A Survey and Reference Design, with a Prototype Interface Case Study
Chest X-ray is the most common radiology exam, yet is limited by the number and variability of radiologists who interpret each image. Deep convolutional neural networks, DenseNet121 in particular, have created a common baseline for Chest X-ray (CXR) interpretation, but accuracy alone cannot define a practically-useful system. We survey the literature for the elements required for a full-stack, explainable CXR interpretation workspace: the state of public data and label quality, DenseNet121 and alternatives, post-hoc explainability (mostly Grad-CAM), in-production architectures, role-based clinical review work-flow, and human-in-the-loop supervision. We describe how we combine these elements into a reference architecture with explainable inference, role-based access to information, logging, and audit trails. We review and compare the capabilities and failure modes of each of the components, review explanation methods per cost and reliability, and detail the role-based access control interface in a prototype workplace MediVision AI. Based on interface captures, the prototype, comprising a multi-label DenseNet121 classifier, Grad-CAM explanations, private workspaces for each user, and one-time expiring share links, is used as a case study to consider trade-offs that emerge in this setting. This demonstration is not a performance study and MediVision AI is not yet field-deployable. We conclude by highlighting open challenges (label noise, distribution shift, subgroup bias, calibration, and regulation) and directions for future research.
Authors
- Hirva Gohel
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24416733
- Primary Topic
- Explainable Artificial Intelligence (XAI)
- Type
- article
- Field-Weighted Citation Impact
- 0.00