A Comparative Study of Transfer Learning Approaches For Medical Image Analysis

Deep learning based approaches have been extensively used for the image classification task with improved accuracy and lower computational complexity. The deep learning approaches have shown outstanding performances when a huge amount of datasets are available to train the model. However, in medical image analysis, sufficient data is not available to train deep learning models. To overcome these issues, a new approach called transfer learning has been introduced for the medical image classification task. This approach has shown outstanding performances with less computational complexity. In the case of transfer learning approaches, the knowledge acquired by training the model with a non-medical image dataset can be transferred to the target dataset. This paper presents a comparative study on the application of transfer learning techniques for the medical image analysis task. This paper demonstrated a complete analysis of the architecture and performance of numerous pre-trained models on the ImageNet dataset. Further, This paper presents the significance of the transfer learning techniques for medical image processing tasks for example brain tumor detection, pneumonia classification, and COVID-19 detection by analyzing various state-of-the-art methods. The experimental results indicate that the performance of transfer learning-based methods can be improved by fine-tuning these models using medical-specific datasets.