Answers are grounded in this paper and cite their sources.
This study evaluates a two-branch deep–texture fusion framework for pediatric pneumonia classification from chest radiographs. EfficientNet-B0 encodes the raw image, while ResNet-18 encodes three gray-level co-occurrence matrix (GLCM) maps: contrast, dissimilarity, and entropy. Squeeze-and-excitation (SE) attention jointly recalibrates the concatenated embeddings before classification. The contribution is this integration strategy and its evaluation, rather than a new backbone, descriptor, or attention operator. The 5,856-image dataset was repartitioned into 4,099 training, 877 validation, and 880 test images, using identical assignments for all configurations. SE fusion achieved accuracy of 0.9818, F1-score of 0.9875, AUROC of 0.9958, and Matthews correlation coefficient of 0.9539. Its accuracy exceeded raw EfficientNet-B0 by five correct predictions and ungated fusion by six. Other configurations achieved higher sensitivity or specificity. The mean normalized texture gate was 0.0759, but this gate summary is not a measure of causal importance; the benefit over simpler auxiliary features remains untested. Attribution and perturbation analyses are model audits, without radiologist-validated localization. The study uses one image-level split, with no external or repeated cross-validation, and patient independence cannot be verified. The results support further investigation on this benchmark but do not establish clinical generalizability or deployment suitability.
No takes yet. Share an insight, caveat, or question.
Kantasa et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: