TL;DR
This paper presents a novel framework for creating a high-resolution, multi-contrast CT atlas of the healthy human pancreas. By employing a deep learning-based preprocessing step and a hierarchical registration pipeline, they successfully align and average CT scans from a large cohort, providing a generalized anatomical reference crucial for understanding pancreatic health and disease.
What Problem Does This Paper Address?
Understanding the human body at a granular level is essential for identifying disease biomarkers and developing effective diagnostic tools. However, significant anatomical variability exists across individuals, particularly in organs like the pancreas, which exhibits notable differences in shape and volume. This variability complicates the generalization of findings from population-wide studies. Existing anatomical atlases often lack the resolution, multi-contrast information, or population-specific representation needed to accurately capture and analyze pancreatic anatomy, hindering the identification of subtle disease-related changes. This paper aims to bridge this gap by creating a standardized, high-resolution atlas of the healthy pancreas across different contrast phases of CT scans.
Methods
The authors developed a multi-stage framework for atlas construction.
- Preprocessing (Body Part Regression - BPR): They utilized a self-supervised deep neural network (adapted from Tang et al.) to standardize the field of view (FOV) across different CT scans. This step is crucial for reducing registration failures caused by variations in imaging protocols and patient anatomy.
- Hierarchical Registration (DEEDS): A two-stage registration process was employed. First, an affine registration aligns the overall anatomy, followed by a dense displacement sampling (DEEDS) deformable registration to refine the alignment and capture fine anatomical details. DEEDS is highlighted for its ability to handle large deformations and its discrete optimization approach, which offers better control over displacement.
- Multi-Organ Segmentation (UNesT): A transformer-based segmentation network (UNesT) was used to generate organ labels for 13 abdominal organs. These labels are vital for evaluating registration accuracy through inverse label transfer and for fusing segmentation results.
- Atlas Construction: After registration, a weighted sliding window averaging algorithm was applied to create the average atlas for each contrast phase. This method helps to reduce blurring and preserve anatomical features.
- Label Fusion: A majority voting approach was used to fuse the segmented pancreas labels from individual subjects onto the atlas, creating a statistically fused pancreas label for visualization and localization.
Datasets
- Clinical Research Multi-Phase CT Cohort: This dataset, sourced from the ImageVU and RD databases, consists of 2000 abdominal CT scans. After excluding patients with pancreatic diseases and applying quality assessment (e.g., pancreas Dice score > 0.1), 443 subjects were used to create the atlas across four contrast phases: non-contrast (59 volumes), arterial (40 volumes), portal venous (330 volumes), and delayed (14 volumes).
- Multi-Organ Labeled Portal Venous Abdominal CT Cohort (BTCV): This dataset comprises 100 de-identified portal venous phase abdominal CT scans, including 20 from the MICCAI 2015 Multi-Atlas Abdomen Labeling challenge. These scans are labeled with 13 abdominal organs, serving as a benchmark for evaluating registration performance.
Key Insights and Ideas
- Importance of Preprocessing for Registration: The study emphasizes that standardizing the FOV using BPR is critical for successful registration, especially when dealing with diverse imaging protocols and patient anatomies. Without it, registration failures are significantly more likely.
- DEEDS Superiority for Abdominal Organs: The paper strongly advocates for the DEEDS registration method, demonstrating its superior performance over ANTS and NIFTYREG for abdominal CT scan registration, particularly for the pancreas. This is attributed to its discrete optimization and dense sampling capabilities.
- Multi-Contrast Atlas Value: Creating separate atlases for different contrast phases is crucial because the pancreas and surrounding organs exhibit distinct appearances and enhancement patterns, which are important for diagnostic purposes.
- Pancreas Stability: Despite overall anatomical variability, the pancreas region showed relatively low variance across registered subjects compared to other abdominal structures like the spleen and spine, suggesting a degree of anatomical consistency that can be leveraged for atlas creation.
Key Results
- High Registration Performance: The two-stage DEEDS registration achieved the highest scores for pancreas label transfer across all phases, with an average Dice score of 0.497-0.505.
- External Validation: External evaluation on 100 portal venous scans showed a mean Dice score of 0.504 for the pancreas, confirming the generalizability of the atlas.
- Comparison of Registration Tools: In a comparison of ANTS, NIFTYREG, and DEEDS, DEEDS consistently outperformed the others in terms of Dice scores for pancreas label transfer, both on the clinical cohort and the labeled BTCV dataset.
- Publicly Available Atlas: The constructed multi-phase pancreas atlases and associated organ labels are made publicly available through the Human Biomolecular Atlas Program (HuBMAP), facilitating further research.
Limitations and Future Work
- Registration Accuracy: While DEEDS shows superior performance, the registration accuracy is still not ideal, with some registration failures and an average Dice score around 0.5 for the pancreas. This could influence the quality of the average atlas.
- Subject Selection: To ensure atlas quality, subjects with very low Dice scores were excluded, reducing the initial cohort size from 898 to 443. This means the atlas might not fully represent the entire population's variability.
- Segmentation Accuracy: The UNesT model, while effective, was trained on portal venous phase data. Segmentation in other phases might be less accurate due to subtle anatomical differences.
- Future Work: The authors suggest exploring more robust registration tools to include more subjects and improve atlas quality. Enhancing segmentation accuracy for the pancreas and improving its localization within the context of pathological changes are also identified as future research directions.
Summary updated