Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MI-SegNet: Mutual Information-Based US Segmentation for Unseen Domain Generalization

Yuan Bi, Zhongliang Jiang, Ricarda Clarenbach, Reza Ghotbi, Angelos Karlas, Nassir Navab

TL;DR

This work tackles the challenge of cross-domain generalization in ultrasound carotid segmentation by explicitly disentangling anatomical (shape) and domain (appearance) features using two encoders and a mutual information loss. The MI-SegNet framework employs data transformations, cross reconstruction, and a reconstruction pathway to encourage informative, decoupled representations, while segmentation relies on the anatomical feature map. Empirical results show superior generalization on unseen US domains without adaptation, and strong performance as a pre-trained model for few-shot domain adaptation, with ablations confirming the benefits of MI regularization and cross reconstruction. The approach offers a practical route toward robust, domain-agnostic US segmentation in real-world clinical settings and can be extended with more realistic US-specific transformations in future work.

Abstract

Generalization capabilities of learning-based medical image segmentation across domains are currently limited by the performance degradation caused by the domain shift, particularly for ultrasound (US) imaging. The quality of US images heavily relies on carefully tuned acoustic parameters, which vary across sonographers, machines, and settings. To improve the generalizability on US images across domains, we propose MI-SegNet, a novel mutual information (MI) based framework to explicitly disentangle the anatomical and domain feature representations; therefore, robust domain-independent segmentation can be expected. Two encoders are employed to extract the relevant features for the disentanglement. The segmentation only uses the anatomical feature map for its prediction. In order to force the encoders to learn meaningful feature representations a cross-reconstruction method is used during training. Transformations, specific to either domain or anatomy are applied to guide the encoders in their respective feature extraction task. Additionally, any MI present in both feature maps is punished to further promote separate feature spaces. We validate the generalizability of the proposed domain-independent segmentation approach on several datasets with varying parameters and machines. Furthermore, we demonstrate the effectiveness of the proposed MI-SegNet serving as a pre-trained model by comparing it with state-of-the-art networks.

MI-SegNet: Mutual Information-Based US Segmentation for Unseen Domain Generalization

TL;DR

This work tackles the challenge of cross-domain generalization in ultrasound carotid segmentation by explicitly disentangling anatomical (shape) and domain (appearance) features using two encoders and a mutual information loss. The MI-SegNet framework employs data transformations, cross reconstruction, and a reconstruction pathway to encourage informative, decoupled representations, while segmentation relies on the anatomical feature map. Empirical results show superior generalization on unseen US domains without adaptation, and strong performance as a pre-trained model for few-shot domain adaptation, with ablations confirming the benefits of MI regularization and cross reconstruction. The approach offers a practical route toward robust, domain-agnostic US segmentation in real-world clinical settings and can be extended with more realistic US-specific transformations in future work.

Abstract

Generalization capabilities of learning-based medical image segmentation across domains are currently limited by the performance degradation caused by the domain shift, particularly for ultrasound (US) imaging. The quality of US images heavily relies on carefully tuned acoustic parameters, which vary across sonographers, machines, and settings. To improve the generalizability on US images across domains, we propose MI-SegNet, a novel mutual information (MI) based framework to explicitly disentangle the anatomical and domain feature representations; therefore, robust domain-independent segmentation can be expected. Two encoders are employed to extract the relevant features for the disentanglement. The segmentation only uses the anatomical feature map for its prediction. In order to force the encoders to learn meaningful feature representations a cross-reconstruction method is used during training. Transformations, specific to either domain or anatomy are applied to guide the encoders in their respective feature extraction task. Additionally, any MI present in both feature maps is punished to further promote separate feature spaces. We validate the generalizability of the proposed domain-independent segmentation approach on several datasets with varying parameters and machines. Furthermore, we demonstrate the effectiveness of the proposed MI-SegNet serving as a pre-trained model by comparing it with state-of-the-art networks.
Paper Structure (17 sections, 7 equations, 2 figures, 2 tables)

This paper contains 17 sections, 7 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Data Augmentation
  3. Image-Level Domain Adaptation
  4. Feature Disentanglement
  5. Method
  6. Mutual Information
  7. Image Segmentation and Reconstruction
  8. Data Transformation
  9. Domain Transformations
  10. Spatial Transformations
  11. Cross Reconstruction
  12. Experiments
  13. Implementation Details
  14. Performance Comparison on Unseen Datasets
  15. Without Adaptation
  16. ...and 2 more sections

Figures (2)

  • Figure 1: Network structure of MI-SegNet. The green and blue arrows represent the data flow of the first ($x_{a_1d_1}$) and the second input image ($x_{a_2d_2}$), respectively.
  • Figure 2: Visual comparison between MI-SegNet and other segmentation networks on US carotid artery datasets without adaptation. For each row, we show the input US image, the ground truth (GT), and the output of each network. Red, pink and green regions represent the false negative, false positive and true positive, respectively.