<![CDATA[We present neural network (NN)–support vector machine (SVM), hybrid NN-SVM framework for three-class pneumonia detection (normal, bacterial, and viral) from chest X-rays (CXRs). Pretrained NN backbone is fine-tuned for radiographic textures; global average pooling (GAP) yields embeddings that feed calibrated radial basis function (RBF)-SVM. Standardized preprocessing (resize, normalization) and class-aware augmentation are applied. We report accuracy, precision, recall, F1-score, area under the curve (AUC), confusion matrices, and per-class receiver operating characteristic (ROC). Statistical significance is assessed via DeLong (AUC), McNemar (accuracy), and paired bootstrap (F1-score). Gradient-weighted class activation mapping (grad-CAM) supports interpretability; external validation and domain adaptation (batch normalization re-estimation and temperature scaling) assess robustness. NN-SVM attains 97.46% accuracy with strong macro-F1 and AUC. Compared with SoftMax head, SVM improves margin separation and calibration. We present NN-SVM, hybrid deep learning approach that combines transfer-learned convolutional neural networks (CNNs) with SVM classifier to automatically diagnose pneumonia from CXRs into three clinically relevant categories: viral pneumonia, bacterial pneumonia, and normal. We use pre-trained CNN to extract robust image embeddings after standardized preprocessing (resizing and normalization) and train RBF-kernel SVM on resulting features. Performance is evaluated with accuracy, precision, recall, F1-score, and confusion matrices. On labeled CXR dataset, NN-SVM achieves 97.46% accuracy, demonstrating strong diagnostic capability that can reduce radiologist burden and support timely clinical decision-making.]]>
Copyrights © 2026
