Abstract:
Medical image classification is not only a complex
task but also a challenging one due to the heterogeneous nature
of medical data. Deep transfer learning has proven to be a viable
technique for medical image classification throughout the years,
mostly because it is able to leverage knowledge from pre-trained
models learned from large-scale datasets, improved
performance, minimal training and overcoming the
disadvantage of small data sets. This paper offers a succinct
review of the cutting-edge deep transfer learning optimization
approaches for medical image classification. The paper begins
with an overview of convolutional neural networks (CNN) and
transfer learning techniques, such as relation-based, feature,
parameter and instance-based transfer learning. Then, the
study examines classical classifiers, such as Resnet, VGG,
Alexnet, Googlenet, and Inception, and compare their
performance on medical image classification tasks. The study
also presents optimization techniques, including batch
normalization, regularization, and weight initialization, data
augmentation and the kernel mathematical formulations.
Finally, the study unearths various challenges that arise when
using deep transfer learning for medical image classification as
well as potential future approaches for this field.
