End-to-end Deep Learning Architectures

© dkfz.de

Current Medical Image Analysis approaches often comprise a set of separate processing steps such as Registration, Normalization, Segmentation, Feature Extraction and Classification. This project develops techniques for the integration of these components into one end-to-end deep learning architecture. This enables simultaneous optimization of all component w.r.t the ultimate clinical task (e.g. disease classification).   

 

 -> Paul Jäger 

Machine Learning-based dMRI Processing

© dkfz.de

This project deals with processing, analysis and visualization of neurological datasets with focus on diffusion-weighted magnetic resonance imaging (dMRI). Major fields of research are the development and implementation of new methods for segmentation or tractography of white matter tracts, as well as tissue segmentation and modelling. Besides the classical methods that are used in this field, we explore the application of machine learning in the context of diffusion-weighted image processing.

 

 

-> Peter Neher

-> Jakob Wasserthal

Validation of Fiber Tractography

© dkfz.de

The quantitative evaluation of fiber tractography is a long-standing challenge for the field that represents an essential prerequisite for widespread application and meaningful interpretation of the approach. In this project we develop phantom-based as well as in-vivo methods to approach this challenge and validate tractography results in large-scale evaluation studies and international challenges.

 

-> Peter Neher

Learning from Noisy Annotations

© dkfz.de

The annotation of medical images suffers from high inter- and intra-rater variability, caused by a number of factors such as strong ambiguities in the images and the subjectivity of annotators. This projects seeks to develop methods that can handle such ambiguous ground truth labels, e.g. by modelling the distribution over annotations via generative adversarial models.
 

-> Simon Kohl

Learning from Weak Annotations

The full annotation of training data with individual labels on each observation can be cumbersome or even impossible in many clinically relevant scenarios. This project aims at developing machine learning methods that can handle weakly annotated data. This can drastically reduce the annotation effort in applications like semantic image segmentation and also enable the application of machine learning in settings where sufficient training data were missing so far.

 

-> Michael Götz

The Black Swan Project

© dkfz.de

This project is part of the “Black Swan Research Initiative”, launched by the International Myeloma Foundation. The goal is to establish an imaging based prognostic staging system for patients with multiple myeloma. In symptomatic patients focal lesions can be present in extensive numbers. The project aims to develop a set of automated detection, segmentation and characterization techniques for whole body image analysis. The system will quantify and assess the tumor mass trend over time.

 

-> Andre Klein

Knowledge-based Large-scale Image Segmentation

© dkfz.de

This project aims at the development of flexible and accurate methods for an automatic semantic image segmentation on large clinical datasets. We focus on a combination of model-based methods with techniques from Active Learning, Online Learning, Transfer Learning and Deep Learning. They allow an optimized training on sparse data, a continuous learning at runtime and an assessment of segmentation quality, as a prerequisite for the successful annotation of large and heterogeneous datasets. 

 

-> Tobias Norajitra

Robust Image-based Stratification of Glioblastoma Patients

© dkfz.de

This project focuses on uncertainty in radiomics. Traditionally, a multitude of parameters is computed from images with which decision support systems are learned. These parameters may however be very sensitive to segmentation errors, resulting in overfitting and degrading performance. We make use of deep learning segmentation algorithms that allow uncertainty estimation, based upon which we can learn more robust decision support algorithms.

 

-> Fabian Isensee

Radiomics

© dkfz.de

In Radiomics, high-throughput computing techniques are systematically employed for the conversion of images to higher-dimensional data, i.e. predictive features. The aim is to improve decision support by the subsequent analysis of these features. We study comprehensive MRI imaging phenotypes for linkage of imaging with clinical, biological and genomic parameters in several entities including prostate cancer, breast cancer and brain tumors.

 

-> Michael Goetz

-> Paul Jäger

-> Tobias Norajitra

 

Quantitative Invasion Prediction and Assessment Support for Glioblastoma Patients

© dkfz.de

In this project, we develop algorithms to predict future clinical assessments of glioblastoma patients from existing longitudinal data. We specifically try to estimate spatial invasion probabilities, incorporating novel uncertainty measures that allow us to quantify the confidence of our models’ outputs. This project is carried out in close cooperation with the Department of Neuroradiology at Heidelberg University Hospital, where we have established infrastructure that allows us to test cutting edge algorithms in clinical routine.

 

 

-> Jens Petersen

to top