Skip to content
Open access

Transfer learning and impact of pretrained deep convolution neural networks on medical image classification: an empirical evaluation with diverse imaging modalities

Nov 2026 · Neural computing & applications (Print) · Vol 38 · 0 citations · 38 references

Abstract

Imaging techniques such as chest X-rays, whole slide images (WSI), and optical coherence tomography (OCT) serve as primary screening and detection tools for a wide variety of pulmonary and ophthalmic conditions. This paper presents a systematic and unified empirical investigation of the impact of ImageNet-pretrained deep convolutional neural networks on medical image classification across three distinct imaging modalities. The study evaluates ten widely adopted CNN architectures under five training strategies: three transfer learning configurations for binary and multiclass classification based on ImageNet pretraining (frozen feature extraction, partial fine-tuning, and full fine-tuning) and two randomly initialized configurations (partial training and full training from scratch). The proposed framework was evaluated on three publicly available medical imaging datasets. For binary classification, the chest X-ray pneumonia dataset consisted of 4,000 training, 1,043 validation, and 624 testing images. For multiclass classification, experiments were performed on the OCT retinal dataset with 86,647 training, 21,662 validation, and 1,000 testing images, and the LC25000 histopathology WSI dataset with 9,600 training, 1,680 validation, and 720 testing images. The model performance was evaluated using AUC, precision, sensitivity, and specificity. All metrics were reported as the mean ± SD along with their corresponding 95% CIs. Our results demonstrate that fixed feature extraction yields substantially lower performance on chest X-ray and OCT, with AUCs falling as low as 0.64–0.88, while histopathology WSI classification benefits notably even under frozen feature extraction, achieving AUCs above 0.96, a performance advantage of approximately 20% over chest X-ray and OCT in this setting. Full fine-tuning consistently improved performance across modalities, with the best chest X-ray AUC reaching 0.9916, the best OCT AUC reaching 0.9701, and the best WSI AUC reaching 0.9980. Contrary to expectation, larger and more complex architectures did not consistently outperform smaller models. These findings indicate that ImageNet performance gains do not translate proportionally to medical imaging tasks, and that the optimal model-modality pairing varies with dataset characteristics, domain shift, and training strategy. This study provides actionable insights for designing transfer learning pipelines in clinical medical image analysis.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.