Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Unified Framework for Foreground and Anonymization Area Segmentation in CT and MRI Data

Michal Nohel, Constantin Ulrich, Jonathan Suprijadi, Tassilo Wald, Klaus Maier-Hein

TL;DR

The paper introduces an open-source preprocessing toolkit for self-supervised learning in 3D medical imaging that tackles two bottlenecks: filtering computationally wasteful air regions via anatomical foreground segmentation and preventing misleading supervision from anonymized data via deface/reface anonymization segmentation. It deploys two nnU-Net-based networks trained on large, multi-modal CT and MRI datasets to delineate foreground anatomy and anonymized facial regions, respectively, with a unified 3D patch approach of $192\times192\times192$ and $z$-score normalization. Empirical results show near-perfect Dice scores ($\approx 99\%$) for foreground segmentation on in-distribution data and robust performance on external datasets, while anonymization segmentation also achieves high Dice across deface/reface procedures, indicating strong applicability for privacy-preserving SSL in CT and MRI. This work enables faster SSL pretraining and safer supervision across diverse 3D medical imaging tasks, potentially accelerating downstream diagnosis and analysis while maintaining data privacy.

Abstract

This study presents an open-source toolkit to address critical challenges in preprocessing data for self-supervised learning (SSL) for 3D medical imaging, focusing on data privacy and computational efficiency. The toolkit comprises two main components: a segmentation network that delineates foreground regions to optimize data sampling and thus reduce training time, and a segmentation network that identifies anonymized regions, preventing erroneous supervision in reconstruction-based SSL methods. Experimental results demonstrate high robustness, with mean Dice scores exceeding 98.5 across all anonymization methods and surpassing 99.5 for foreground segmentation tasks, highlighting the efficacy of the toolkit in supporting SSL applications in 3D medical imaging for both CT and MRI images. The weights and code is available at https://github.com/MIC-DKFZ/Foreground-and-Anonymization-Area-Segmentation.

A Unified Framework for Foreground and Anonymization Area Segmentation in CT and MRI Data

TL;DR

The paper introduces an open-source preprocessing toolkit for self-supervised learning in 3D medical imaging that tackles two bottlenecks: filtering computationally wasteful air regions via anatomical foreground segmentation and preventing misleading supervision from anonymized data via deface/reface anonymization segmentation. It deploys two nnU-Net-based networks trained on large, multi-modal CT and MRI datasets to delineate foreground anatomy and anonymized facial regions, respectively, with a unified 3D patch approach of and -score normalization. Empirical results show near-perfect Dice scores () for foreground segmentation on in-distribution data and robust performance on external datasets, while anonymization segmentation also achieves high Dice across deface/reface procedures, indicating strong applicability for privacy-preserving SSL in CT and MRI. This work enables faster SSL pretraining and safer supervision across diverse 3D medical imaging tasks, potentially accelerating downstream diagnosis and analysis while maintaining data privacy.

Abstract

This study presents an open-source toolkit to address critical challenges in preprocessing data for self-supervised learning (SSL) for 3D medical imaging, focusing on data privacy and computational efficiency. The toolkit comprises two main components: a segmentation network that delineates foreground regions to optimize data sampling and thus reduce training time, and a segmentation network that identifies anonymized regions, preventing erroneous supervision in reconstruction-based SSL methods. Experimental results demonstrate high robustness, with mean Dice scores exceeding 98.5 across all anonymization methods and surpassing 99.5 for foreground segmentation tasks, highlighting the efficacy of the toolkit in supporting SSL applications in 3D medical imaging for both CT and MRI images. The weights and code is available at https://github.com/MIC-DKFZ/Foreground-and-Anonymization-Area-Segmentation.
Paper Structure (8 sections, 3 figures, 2 tables)

This paper contains 8 sections, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Materials and methods
  3. Anatomical foreground segmentation
  4. Deface area segmentation
  5. Results
  6. Anatomical foreground segmentation
  7. Deface area segmentation
  8. Discussion

Figures (3)

  • Figure 1: Boxplot results illustrating the Dice Similarity Coefficient and HD95 for foreground segmentation, shown for both the test portion of the training dataset and the external test dataset.
  • Figure 2: Boxplot examples illustrating the Dice coefficient and HD95 for segmentation of deface area across different anonymization modes and the external OpenNeuro dataset.
  • Figure 3: Example predictions for foreground (FG) segmentation and anonymization segmentation. While in the majority of cases, the FG segmentation is almost perfectly solved (a), we observed a rare failure case for volumes with an artificial boundary between actual air and a region with constant values introduced during image reconstruction(b). Further we observed problems in images without a sharp contrast between the background and surrounding body regions, where even a human would not be able to deliniate the background (c). Lastly, the ground truth masks for the Deface anonymization (green) in the Open Neuro dataset were occasionally ambiguous. While the network accurately predicted the removed portions of the head in the foreground (d-e), we observed in some cases an ambiguity of the anonymization ground truth mask in the background (f).