MaNOISE

Chair for Computer Aided Medical Procedures & Augmented Reality
Lehrstuhl für Informatikanwendungen in der Medizin & Augmented Reality

Robust inference: How to handle noisy labels in medical applications?

Supervision: Prof. Dr. Nassir Navab, Dr. Seong Tae Kim

Abstract

The performance of supervised learning methods highly depends on the quality of the labels. In some medical applications, there are ambiguities to accurately assign a class label from the image alone. In a typical labeling process, multiple medical experts provide subjective annotations which are also affected by their varying skill-levels and biases. For these reasons, the physicians often fail to make consensus on their annotations regarding the presence or location of the disease. Simply treating these noisy labels as the ground truth limits the accuracy of learning algorithms, in particular, when there is strong disagreement. To overcome the aforementioned limitations, in this project, we try to develop a new learning scheme to handle the noisy label.

Requirements:

  • Good understanding of statistics and machine learning methods.
  • Very good programming skills in Python & TensorFlow? / PyTorch?

Location:

  • Garching

References:

[1] S. Kohl et al. “A Probabilistic U-Net for segmentation of ambiguous images,” NeurIPS? 2018.

[2] K. Lee et al. “Robust inference via generative classifiers for handling noisy labels,” ICML 2019.

[3] R. Tanno et al. “Learning From Noisy Labels By Regularized Estimation Of Annotator Confusion,” CVPR 2019.

If you are interested, please contact us via e-mail: seongtae.kim@tum.de

ProjectForm
Title: Robust inference: How to handle noisy labels in medical applications?
Abstract: The performance of supervised learning methods highly depends on the quality of the labels. In some medical applications, there are ambiguities to accurately assign a class label from the image alone. In a typical labeling process, multiple medical experts provide subjective annotations which are also affected by their varying skill-levels and biases. For these reasons, the physicians often fail to make consensus on their annotations regarding the presence or location of the disease. Simply treating these noisy labels as the ground truth limits the accuracy of learning algorithms, in particular, when there is strong disagreement. To overcome the abovementioned limitations, in this project, we try to develop a new learning scheme to handle the noisy label.
Student:  
Director: Prof. Dr. Nassir Navab
Supervisor: Dr. Seong Tae Kim
Type: Master Thesis
Area: Segmentation, Machine Learning, Medical Imaging, Computer Vision
Status: open
Start:  
Finish:  
Thesis (optional):  
Picture:  


Edit | Attach | Refresh | Diffs | More | Revision r1.2 - 02 Aug 2019 - 09:51 - SeongTaeKim