Multi-scale Feature Learning of Domain-robust Pneumonia Detection and Cross-regional Anatomical Attention Cross-dataset Validation Study

Most existing deep learning-based systems to detect pneumonia in chest radiography have no anatomical awareness and have poor cross-dataset generalizability, limiting their use in clinical environments.Our model is an Anatomically Guided Multi-Scale Cross-Regional Attention Network (AGMSCRA-Net), which contains: (1) crossregional attention networks that model attentional relationships between seven anatomical lung regions, (2) a multiscale feature pyramid that models infiltrates at different resolutions, (3) a four-phase curriculum learning algorithm that proceeds through easy to hard samples, and (4) anatomy consistency loss that enforces clinical consistency in prediction.The model was trained and evaluated on two independent datasets: Chest X-Ray Pneumonia (5,216 training samples and 624 test samples) and COVID-19 & Pneumonia (4,684 training samples and 1,288 test samples) datasets.AGMSCRA-Net achieved same-domain accuracies of 97.360% and 97.520% on the pneumonia and COVID-19 datasets, respectively, and cross-domain accuracies of 97.440% and 97.120%, respectively.There was high crossdataset generalization, and the average difference in performance between exactly equal domains was minimal (0.160% worse in cross-domain than same-domain: 97.440% versus 97.280%, respectively).Ablation studies found that architecture is resistant to removing individual parts, whose removal alone has a minimal effect on accuracy.This work introduces an anatomically aware deep learning model that significantly improves cross-dataset generalizability in pneumonia detection on chest radiographs.Therefore, providing clinicians and healthcare organizations with a more reliable, clinically resilient artificial intelligence system for implementation in clinical practice.