Improved Medical Imaging Transfer Learning through the Conflation of Domain Features

Transfer learning has made deep learning more accessible in many fields, such as medical imaging. However, data adaptation in medical imaging transfer learning remains a challenge. With the release of many pre-trained models, there is a need to address target data adaptation in these pre-trained modes. This paper proposes the use of conflation of textural features, testing it on three medical imaging datasets and two pre-trained models, among them a MobileNetV2, to demonstrate the approach’s usefulness in mobile systems. From the experiments, the selection of images with lower textural Kullback-Leibler divergence is seen to improve the performance accuracy of the models by a margin of 13.17% in LBP and 6.47% for GLCM methods. This approach ensures that the pre-trained models can be used with much confidence and assist in generating more quality data samples for effective transfer learning in medical imaging and other applications using image data.