Distributed system for the analysis of medical x-ray images using neural networks

This paper presents the development, calibration, and deployment of a system for the automated diagnosis of pneumonia based on chest X-ray images. Addressing the limitations of standard architectures trained on non-medical data, the domain-specific BioViL model was selected as the backbone, constituting a key aspect of the study's novelty. A complete and reproducible pipeline is described, ranging from data preprocessing utilizing lung segmentation to multi-stage calibration for handling imbalanced datasets, and finally, the deployment of the solution as a distributed client-server application. The system achieved an overall accuracy of 88.9% and an AUC-ROC score of 0.929. The model demonstrated an excellent balance, exhibiting high sensitivity for pneumonia (93.6%) and a specificity of 81.2%, confirming the effectiveness of the developed balanced solution.