Pneumonia Detection Based on Deep Transfer Features and Heterogeneous Ensemble Learning
Jiahuan Li, Xueri Li, Lei Yang, Shan Peng, Bingbing Liao, Wenbo Xie, Dheeraj Kumar
The rapid advancement of traditional machine learning has opened up new avenues within the medical field. For the automated screening of paediatric pneumonia via chest radiographs, this paper proposes a ‘deep features + heterogeneous ensemble’ framework. Utilising Kaggle’s dataset of 5,856 AP-position chest radiographs from children aged 1-5 years (with a pneumonia-to-normal ratio of approximately 3:1), images were uniformly resized to 224×224 pixels. Features were extracted using a pre-trained ResNet-50 model from ImageNet, yielding 2,048-dimensional representations. A PCA-based dimensionality reduction method retaining 95% variance served as a control. Two sets of base learners were designed: A: Random Forest + XGB + Support Vector Machine; B: Logistic Regression + Gradient Boosting Decision Tree + LightGBM. These output positive class probabilities, which were then fused by the CatBoost meta-learner after passing through a StandardScaler. Experiments demonstrated: Group A achieved pneumonia F1 0.92 and recall 0.97 on raw features, but normal recall was only 0.76; after PCA, Group B elevated normal recall from 0.65 to 0.83, while pneumonia precision improved from 0.82 to 0.90. The composite F1 reached 0.85/0.91, achieving the most balanced inter-class performance with the lowest false positives. ROC curves indicate that Group B’s PCA approach maintains an AUC exceeding 0.95. These findings confirm that judiciously combining base learners with dimensionality reduction strategies can significantly reduce false positives in healthy children while maintaining high pneumonia detection rates. This provides a clinically viable paradigm balancing accuracy, recall, and robustness.