Introduction
Self-supervised learning (SSL) has emerged as a powerful paradigm for leveraging large-scale unannotated medical imaging data. However, applying SSL to 3D CT and MRI scans presents two significant challenges: computational inefficiency and the presence of anonymization artifacts. 3D volumes often contain substantial "air" or background voxels that lack anatomical information, leading to wasted GPU cycles during random patch sampling. Furthermore, common anonymization techniques—such as blurring facial features—introduce artificial noise. For reconstruction-based SSL methods, these artifacts provide erroneous supervision, potentially degrading the quality of learned features.
Methodology
To address these hurdles, we propose a unified framework utilizing the nnU-Net architecture to develop two specialized segmentation networks:
- Foreground Segmentation Network: Designed to delineate all anatomical regions, enabling "informed sampling" where models focus exclusively on relevant tissues.
- Anonymization Area Segmentation Network: Specifically targets regions modified by tools like AFNI’s deface and reface. This allows SSL loss functions to mask out these areas, ensuring models learn only from authentic anatomy.
The framework achieves cross-modality robustness by employing z-score normalization and a large patch size ($192 \times 192 \times 192$), allowing a single configuration to process both CT and MRI data effectively.
Datasets and Validation
The models were trained on a diverse collection of 3,299 3D images (1,899 CT, 1,400 MRI) sourced from multiple public repositories, including LACS, AMOS, and TotalSegmentator. Anonymization training utilized the OASIS3 dataset, processed with three distinct AFNI methods. External validation was performed on the OpenNeuro dataset to assess generalizability to unseen data distributions.
Results
The framework demonstrated exceptional performance across both tasks. The foreground segmentation model achieved a mean Dice Similarity Coefficient (DSC) of 99.56% on internal tests and 98.57% on external data.
Similarly, the anonymization detection network proved robust across various defacing schemes, maintaining mean Dice scores above 98.5%.
Discussion and Limitations
While the framework is highly effective, certain edge cases persist. In CT imaging, sharp boundaries introduced by reconstruction algorithms can sometimes be misinterpreted as foreground. Additionally, low-contrast boundaries in specific MRI sequences can lead to minor precision losses. Future work will focus on refining the models to handle these reconstruction-specific artifacts and expanding support for a wider range of anonymization techniques. By providing a robust, open-source toolkit for these preprocessing steps, we aim to streamline the development of high-quality SSL models for 3D medical imaging.
Summary updated