EE Seminar: Effectively Optimizing Medical Transfer Learning using Colormaps
Speaker: Michael Zolotov
M.Sc. student under the supervision of Prof. Hayit Greenspan
Wednesday, December 19th, 2018 at 15:00
Room 011, Kitot Bldg., Faculty of Engineering
Effectively Optimizing Medical Transfer Learning using Colormaps
Abstract—The medical world is based today on computed tomography (CT) imaging technology to detect pulmonary cancer. This process suffers from countless significant bottlenecks and difficulties: from acute dependence on long and meaningful training for the diagnoses of CT scans, to human errors resulting from fatigue and lack of concentration.
In recent years, there has been a technological breakthrough in machine learning, especially in systems based on deep neural networks. The performance of neural network based systems in detecting and classifying real-world images is state-of-the-art, clearly bypassing the performance of older algorithms.
Transfer learning is key principle for building high performance object detection systems. It is a method where a model which was developed for some task is reused as the starting point for a second task.
Transfer learning is useful as it enables training deep neural networks with comparatively little data. However, because medical images are different from ordinary images, the effect of applying naive transfer learning techniques is far weaker than usual. Therefore, current models that are applied for Computer Aided Diagnosis (CAD) tasks do not reach their full potential.
This work shows that by using colormaps, it is possible to optimize transfer learning for medical images.
Colormaps are used nowadays to find a visually comfortable data representation in 3D colorspace. They are used in this work as a pre-processing technique, for the first time, to the best of our knowledge.
We will show that by using colormaps it is possible to improve the overall detection performance of Deep Learning based systems on medical images, when fine tuning models that have been pretrained on ImageNet, without changing the detection architecture itself.
