Channel attention recalibrated fusion of CoAtNet and frozen DINOv2 features for binary chest X ray classification

Abstract CXRNet is proposed for binary chest X ray classification of NORMAL versus PNEUMONIA using the Chest X Ray Images Pneumonia dataset, derived from the pediatric cohort reported in Cell 2018. The dataset includes 5,856 images (NORMAL 1,583; PNEUMONIA 4,273) with a held out test set of 624 images (NORMAL 234; PNEUMONIA 390). A leakage safe evaluation protocol is applied through strict split isolation, exact duplicate and near duplicate screening prior to training, and training only augmentation, with no augmentation applied to validation or test data. CXRNet follows a dual branch attention guided fusion design that combines a CoAtNet feature extractor for hierarchical local to global representations with a frozen pretrained DINOv2 encoder used strictly as a global feature extractor. The branch features are aligned, concatenated, and recalibrated using a channel attention module before a softmax classifier. The model is implemented in TensorFlow Keras on Windows 11 with an NVIDIA RTX 3090 GPU and trained using Adam with learning rate 0.0001, batch size 32, and 100 epochs. CXRNet achieves 98.26% training accuracy and 97.92% test accuracy, with class-wise performance of 95.85% precision, 98.72% recall, and 97.26% F1 for NORMAL, and 99.22% precision, 97.44% recall, and 98.32% F1 for PNEUMONIA; macro averaged precision, recall, and F1 are 97.53%, 98.08%, and 97.79%, while weighted averages are 97.95%, 97.92%, and 97.92%. Grad-CAM visualizations are provided for qualitative interpretability, and comparisons with Swin Transformer, ConvNeXt Tiny, ResNet50 plus InceptionV3, and NASNetMobile under identical training settings show that CXRNet yields the strongest test accuracy among the evaluated models on this dataset.