Transfer Learning Architectures for Chest Radiograph Classification: A Systematic Review and Meta-Analysis of Accuracy, Model Scale and Data Efficiency

Abstract

Transfer learning from ImageNet-pretrained convolutional and transformer architectures is the default approach to chest radiograph classification, and the literature reports accuracies above 98% with some regularity. Whether those figures reflect genuine diagnostic capability, and whether larger architectures deliver better radiographic performance, has not been established quantitatively. Objective. To pool reported performance across transfer-learning studies of chest radiograph classification, quantify heterogeneity and small-study effects, and test whether model scale predicts diagnostic performance. Methods. Structured searches across six sources retrieved 94 records; 31 studies met eligibility and were assigned to three strata: binary paediatric pneumonia classification (k = 13), multi-label ChestX-ray14 AUROC (k = 20), and a controlled set of 12 architectures evaluated under an identical protocol. Accuracies were pooled on the logit scale using DerSimonian–Laird random-effects models; heterogeneity was quantified with Q, τ² and I²; small-study effects were assessed by Egger's regression; and model scale was tested by meta-regression on log₁₀ parameters. Results. Pooled binary accuracy was 96.22% (95% CI 94.42–97.46%) but with extreme heterogeneity (Q = 272.0, df = 12, I² = 95.6%, τ² = 0.501, p < 0.001), and Egger's test indicated strong small-study effects (intercept 5.42, SE 2.19, p = 0.0001). On ChestX-ray14, transformers did not outperform convolutional networks (0.8121 vs 0.8018 mean AUROC; Welch t = −0.814, p = 0.440), while hybrid architectures led at 0.8344. Model scale was negatively associated with performance: meta-regression on ChestX-ray14 gave a slope of −0.023 AUROC per decade of parameters (r = −0.594, p = 0.054), and in the controlled set, excluding two models that failed to converge, F1 fell 2.82 points per decade of parameters (r = −0.691, p = 0.027). MobileNetV3 (3M parameters) achieved the highest F1 of all twelve architectures, exceeding VGG-16 (135M) by 9.0 points. Conclusions. The headline accuracies in this literature are not a stable estimate of anything: heterogeneity is near-total and publication bias is strong. Where architectures are compared under identical conditions, small models win a finding with direct implications for deployment in the resource-constrained settings where chest radiography matters most.