Bayeisan One-Encoder-Multiple-Decoder (BOEMD) Network

Qingqiao Hu, Hao Wang, Jing Luo, Yunhao Luo, Zhiheng Zhang, Jan S. Kirschke, Benedikt Wiestler, Bjoern Menze, Jianguo Zhang, Hongwei Bran Li
Jun 20, 2023
PDF Code

In this project, we1 proposed a simple yet effective Bayesian Neural Network (BNN) framework, especifically designed for inter-rater uncertainty quantification probelm in medical imaging. We call this framework, Bayesian One-Encoder-Multiple-Decoder (BOEMD) network.

A qualitative comparison was performed among the probabilistic U-Net, PHiSeg and BOEMD. The error map quantifies the difference between predicted and observed distributions. From the segmentation outputs and the error map, one can see that our method captures inter-rater uncertainty more accurately compared to the others.
A qualitative comparison was performed among the probabilistic U-Net, PHiSeg and BOEMD. The error map quantifies the difference between predicted and observed distributions. From the segmentation outputs and the error map, one can see that our method captures inter-rater uncertainty more accurately compared to the others.

Automated medical image segmentation inherently involves a certain degree of uncerainty. One key factor contributing to this uncertainty is the ambiguity that can arise in determining the boundaries of a target region of interest, primarily due to variations in image appearance. On top of this, even among experts in the field, different opinions can emerge regarding the precise definition of specific anatomical structures. Such modeling of segmentation uncertainty, known as inter-rater uncertainty is explored and analyzed in this project.

A schematic view of our architecture to estimate inter-rater uncertainty with Bayesian modeling. It contains one Bayesian encoder and multiple Bayesian decoders. In the skip connection (red, blue, and green) between the encoder and each decoder, an attention module (denoted by A) is introduced to capture rater-specific representation for individual raters.
A schematic view of our architecture to estimate inter-rater uncertainty with Bayesian modeling. It contains one Bayesian encoder and multiple Bayesian decoders. In the skip connection (red, blue, and green) between the encoder and each decoder, an attention module (denoted by A) is introduced to capture rater-specific representation for individual raters.

BOEMD has three main features:

  • One-Encoder-Multiple-Decoder: the only one encoder is designed to learn the shared knowledge from segmentation while the multiple decoders that are aligned with each rater (if the rater ids are known, the encoder tends to learn the specific features from each rater; otherwise, the rater ids are unknown, each decoder learns the shared knowledge as the encoder does.)
  • Rater-specific Attention: each attention unit is associated with one decoder, to further enhance the ability of decoder capturing the specific knowledge from segmentation map.
  • Bayesian Modelling: usually medical imaging dataset is quite small compared to other existing vision dataset, which results in the neural network over-fitting. Adopting Bayesian Modelling in our framework helps the neural network learn the dataset distribution from a high level. Doing this also reduces over-fitting.

We also introduced a rater-aligned liver tumor dataset in this work to further demonstrate the intention of our design.

  1. Qingqiao Hu and Hao Wang made equal contribution to this project. ↩︎

Uncertainty quantification Segmentation Inter-rater variability Deep Learning

Publications

Inter-Rater Uncertainty Quantification in Medical Image Segmentation via Rater-Specific Bayesian Neural Networks
Automated medical image segmentation inherently involves a certain degree of uncertainty. One key factor contributing to this …
Qingqiao Hu, Hao Wang, Jing Luo, Yunhao Luo, Zhiheng Zhangg, Jan S. Kirschke, Benedikt Wiestler, Bjoern Menze, Jianguo Zhang, Hongwei Bran Li
PDF Code Project