/CNN_for_Multiple_Sclerosis

Project done under the direction of David Atienza at the Embedded System Laboratory at EPFL, Lausanne, Switzerland.

Primary LanguageJupyter Notebook

CNNs for Multiple Sclerosis lesion detection

Drawing Drawing

Project carried out within the framework of the European project DeepHealth using the infrastructure of the EPFL Embedded Systems Laboratory. The final report can be accessed here.

  • Author: Francisco Javier Blázquez Martínez
  • Director: Prof. David Atienza Alonso
  • Supervisors: Dr. Arman Iranfar, Dr. Marina Zapater, Dr. Tomás Teijeiro Campo

Abstract

Deep Learning techniques applied to medicine have proven to be more and more useful in recent years. The DeepHealth emerges as a way to provide state-of-the-art deep learning high performance computing tools adapted to their usage in medicine. In this project we create 2D and 3D models for detecting Multiple Sclerosis lesions in Magnetic Resonance Images and we evaluate the performance of EDDL library and compare it against tensorflow.

Project structure

In this project we first created, evaluated and profiled 2D image segmentation models starting with the classical U-Net. EDDL and tensorflow models can be accessed here. Their structure schema is the following:
U-Net

The other 2D model analyzed was a later improvement over the U-Net model, the Double U-Net. EDDL and tensorflow models can be accessed here. Double U-Net architecture is the following:
Double U-Net

We finally analyzed a 3D cascade model proposed by Sergi Valverde with the following structure:
3D cascade model