DOI: 10.70562/tubid.2000572 ISSN: 2528-8652

Multidimensional Performance and Model Selection Analysis of Transfer Learning-Based Deep Learning Models on Multiple Medical Imaging Datasets

Lokman Doğan
Show PDF Cite
Transfer learning-based deep learning models have been widely used in recent years because they can achieve high classification accuracy with a limited amount of labeled medical image data. However, there are few comparative studies that jointly evaluate the accuracy, computational efficiency, and performance consistency across different medical image datasets of current transfer learning models. In this study, the performance of the ResNet50, DenseNet121, EfficientNetB0, EfficientNetV2B0, MobileNetV3, and ConvNeXt-Tiny models, using a transfer learning approach, was comprehensively compared across the PneumoniaMNIST, DermaMNIST, and BloodMNIST datasets. All models were trained on Google Colab using an NVIDIA A100-SXM4-80GB GPU, with the same training parameters, the same data splits, and the same experimental environment. Training utilized the AdamW optimization algorithm, a learning rate of 1×10⁻⁴, a batch size of 32, a maximum of 10 epochs, an early stopping mechanism, and a fixed random seed of 42. The models were evaluated in terms of accuracy, precision, recall, macro F1-score, ROC-AUC, training time, inference time, frames per second (FPS), model size, and number of parameters. The experimental results showed that model performance varied depending on the dataset; all models achieved very high classification accuracy on the BloodMNIST dataset, while the DermaMNIST dataset proved to be the most challenging for the models. In the overall evaluation, the DenseNet121 and ConvNeXt-Tiny models demonstrated the highest average classification performance, while the MobileNetV3 and EfficientNetB0 models stood out in terms of computational efficiency with their smaller model sizes and faster inference times. The findings provide guidance to practitioners and researchers on selecting the appropriate transfer learning model for various medical image classification problems.

More from our Archive