Project scope

DeepAnatomy is a student project that deals with Deep Learning in medical imaging. It has been carried out since the winter semester of 2019 in cooperation with the University of Bremen and the Fraunhofer Institute for Digital Medicine MEVIS.
At the beginning of the project, the project participants acquire general and medical-specific Deep Learning methodologies and learn how to handle the software and hardware resources used by the Fraunhofer Institute for Digital Medicine MEVIS. Subsequently, the project participants focus on identifying and implementing specific sub-projects, thus making a contribution to a highly topical research area.

Main focus 2023/2024

After the project participants had acquired the basic - theoretical as well as practical - skills, they decided on the following sub-projects:
  • Training of different classifiers for the detection of abdominal and paravertebral muscles
  • Sparse Label Annotation via extended SATORI Tool
  • Research about evaluation metrics for medical image segmentation
  • Several infrastructure topics
    • Blossom: Reboot and improvement of the live lossplot tool
    • Keras Migration to Version 3
    • Extension of the Satori file upload functionality
    • Implementation of a OpenSearch based textual logger connected to various cloud based services

With kind permission from MeVis Medical Solutions AG

Deep Learning in Medical Imaging

The analysis of medical imaging data, such as computed tomography or magnetic resonance imaging scans, is a central component of various diagnostic and therapeutic procedures. Examples include early detection of tumors, surgical planning, minimally invasive procedures, and monitoring of therapy progress. The development of Deep Learning algorithms for specific questions of medical image processing and suitable software applications can support doctors in exploiting the full potential of modern imaging procedures and use it for the benefit of patients. This is exactly where DeepAnatomy comes in.

© Haslob Kruse + Partner

Infrastructure

The project participants have the following, special software resources available through the Fraunhofer Institute for Digital Medicine MEVIS:

  • MeVisLab - a framework for medical image processing and visualization in research and development.
  • RedLeaf - a Deep Learning Framework developed by the Fraunhofer Institute for Digital Medicine MEVIS. RedLeaf builds on widely used Deep Learning Frameworks such as Keras and PyTorch. In addition to the integration of MeVisLab, RedLeaf offers numerous predefined neural network architectures and live monitoring of training performance.
  • Challengr - a web interface developed by the Fraunhofer Institute for Digital Medicine MEVIS for comparing different neural networks.
For training neural networks, the project participants have a computer cluster consisting of nine computers, each with four GTX 1080 Ti graphics cards available.

Prior Knowledge and Requirements

Initial knowledge in Deep Learning facilitates project entry, but is not absolutely necessary due to the intensive theoretical training. Prior knowledge in the following areas is also helpful:

Python

Python is a programming language with clear syntax and easy readability, which can be used versatilely. In the project, the Deep Learning Frameworks Tensorflow/Keras and PyTorch as well as the WebFramework Flask are also used.

Docker

Docker is freely available software that simplifies the deployment of applications: the applications are packaged into containers that contain all the packages needed for execution (libraries, system tools, environment variables, etc.).

Quasar

Quasar is a freely available, Vue.js-based framework for frontend development.

About us

The Bachelor project Team 2023/2024

Tamari Bayer

Emrullah Baykara

Felix Drees

Lukas Garbade

Arbresh Gashi

Hanno Henke

Ardit Keta

Jonathan Kinkel

Hannah Koeper

Tobias Liese

Ole Mahlstaedt

Semjon Nirmann

Jorma Reiners

Jessica Repty

Freja Sender

Jörn Steffens

Aziz Tas

Tom Wolff

Lecturers

Prof. Dr. Horst Hahn

Dr. Hans Meine

Felix Thielke