From data to diagnosis: evaluating the role of data-centric techniques in pneumonia segmentation with privacy considerations

Accurate segmentation of pneumonia in chest X-ray (CXR) images is challenging due to noise, variability in image quality, and limited annotated data. This study evaluates the impact of data-centric techniques, including preprocessing and augmentation, on improving segmentation performance. Experiments were conducted on a publicly available dataset under two conditions: without and with data-centric enhancements. Performance was assessed using accuracy, precision, recall, F1-score, Dice Similarity Coefficient (DSC), and Intersection over Union (IoU). Results show significant improvement, with accuracy increasing from 69 to 85%, recall from 50 to 89%, DSC from 62 to 80%, and IoU from 55 to 81%. However, statistical significance testing was not performed and remains future work. In addition, the proposed data-centric pipeline is designed to support privacy-aware learning environments through decentralized preprocessing compatibility. These findings highlight the importance of data-centric strategies, although further validation is required before clinical application.