Understanding the Mechanisms of Deep Transfer Learning for Medical Images
This addresses the challenge of limited training data in medical imaging for researchers and practitioners, but it is incremental as it builds on existing transfer learning methods.
The paper tackles the problem of understanding how deep transfer learning works for medical images, specifically by transferring a CNN trained on ImageNet to kidney detection in ultrasound images, and shows that a transferred and tuned CNN outperforms a state-of-the-art feature engineered pipeline, with a hybridization achieving 20% higher performance.
The ability to automatically learn task specific feature representations has led to a huge success of deep learning methods. When large training data is scarce, such as in medical imaging problems, transfer learning has been very effective. In this paper, we systematically investigate the process of transferring a Convolutional Neural Network, trained on ImageNet images to perform image classification, to kidney detection problem in ultrasound images. We study how the detection performance depends on the extent of transfer. We show that a transferred and tuned CNN can outperform a state-of-the-art feature engineered pipeline and a hybridization of these two techniques achieves 20\% higher performance. We also investigate how the evolution of intermediate response images from our network. Finally, we compare these responses to state-of-the-art image processing filters in order to gain greater insight into how transfer learning is able to effectively manage widely varying imaging regimes.