A systematic benchmarking of deep transfer learning architectures and optimization strategies for COVID-19 detection from lung ultrasound images

Introduction The COVID-19 pandemic imposed severe burdens on global health infrastructure, underscoring the urgent need for rapid, reliable, and scalable diagnostic technologies. Conventional radiological modalities such as chest X-rays and computed tomography (CT) face critical deployment barriers, including ionizing radiation exposure, elevated procedural costs, and substantial infection-control risks in resource-constrained settings. Lung ultrasound (LUS) presents a compelling alternative owing to its portability, absence of ionizing radiation, and suitability for point-of-care application; however, systematic analysis of deep learning architectures optimized for LUS-based COVID-19 classification remains limited. Methods To address this gap, this study presents a comprehensive benchmarking framework in which 13 pretrained convolutional neural network (CNN) architectures—AlexNet, VGG-16, VGG-19, ResNet-50, ResNet-101, ResNet-152, DenseNet-121, DenseNet-169, DenseNet-201, InceptionV3, Inception-ResNetV2, MobileNetV2, and Xception—were evaluated under seven gradient-based optimization strategies: SGD, Adagrad, RMSProp, Adadelta, Adam, AdaMax, and Nadam. Transfer learning was applied uniformly across all architectures to compensate for the limited scale of the publicly available POCOVID-Net dataset, which comprised 455 LUS frames from confirmed COVID-19 cases and 226 frames from non-COVID subjects. In addition, geometric transformation-based data augmentation was employed to mitigate the effects of class imbalance and improve inter-patient generalization. Results Among the architecture–optimizer combinations evaluated under the hold-out scenario, DenseNet-121 with the Adam optimizer achieved perfect classification with 100% accuracy, surpassing all previously reported LUS-based COVID-19 detection methods evaluated under comparable single-dataset conditions. Given the risk that a single hold-out split may overstate performance, rigorous stratified 5-fold cross-validation was subsequently performed on this best-performing hold-out model to ensure robustness and address concerns regarding overfitting. This yielded a mean accuracy of 99.11% ± 0.35%, sensitivity of 99.41% ± 0.31%, specificity of 99.18% ± 0.38%, F1-score of 99.36% ± 0.29%, and AUC of 99.45% ± 0.25%. Discussion/Conclusion Despite consistently strong performance across both evaluation schemes, these results were derived from a single, limited-diversity public dataset and, therefore, represent a methodological benchmark rather than a validated estimate of real-world performance. External multicenter validation is necessary before clinical application of the model. Nonetheless, the evaluated DenseNet-121 framework holds practical utility as a rapid triage aid and complementary screening tool within clinical workflows, particularly in low-resource environments where CT infrastructure is unavailable.