AATCT-IDS: A benchmark Abdominal Adipose Tissue CT Image Dataset for image denoising, semantic segmentation, and radiomics evaluation

Zhiyu Ma, Chen Li, Tianming Du, Le Zhang, Dechao Tang, Deguo Ma, Shanchuan Huang, Yan Liu, Yihao Sun, Zhihao Chen, Jin Yuan, Qianqing Nie, Marcin Grzegorzek, Hongzan Sun

TL;DR

The researchers have released AATCT-IDS, a comprehensive, open-source benchmark dataset of 13,732 abdominal CT slices from 300 subjects. This dataset is specifically designed to advance computer-aided detection of adipose (fat) tissue, providing high-quality expert annotations to evaluate image denoising, semantic segmentation, and radiomics.

What Problem Does This Paper Address?

Metabolic Syndrome (MetS) and obesity are surging global health crises, closely linked to cardiovascular disease. While clinical metrics like Body Mass Index (BMI) and waist circumference are standard, they are "blunt instruments"—they cannot accurately distinguish between subcutaneous adipose tissue (SAT) (fat under the skin) and visceral adipose tissue (VAT) (fat surrounding internal organs), the latter of which is a much higher risk factor for disease.

Computed Tomography (CT) is the gold standard for visualizing these distributions, but the field has been hampered by a "data desert." Most existing datasets are small, private, or lack the rigorous manual annotations needed to train deep learning models. The authors created AATCT-IDS to provide a standardized, large-scale playground for researchers to develop better diagnostic tools.

Methods

The authors didn't just release data; they performed an exhaustive baseline evaluation across three domains:

  1. Image Denoising: They tested seven classical and transform-domain algorithms (including BM3D, Wavelet Transforms, and Non-local Means) against various noise types (Poisson, Gaussian, Speckle, and Pulse) to see which best preserved the delicate boundaries of fat tissue.
  2. Semantic Segmentation: They trained and compared eight deep learning architectures—ranging from the classic U-Net and FCN8s to more modern, lightweight models like BiSeNet and ICNet—to automatically map SAT and VAT regions.
  3. Radiomics & Clustering: They extracted multidimensional features (texture, area, perimeter) and used unsupervised learning (K-Means, DBSCAN, and AGNES) to see if the data could naturally reveal different "phenotypes" of fat distribution in the population.

Datasets

The core contribution is the AATCT-IDS dataset, sourced from Shengjing Hospital of China Medical University.

  • Scale: 300 subjects, 13,732 raw CT slices.
  • Annotations: 3,213 slices were manually annotated by seven biomedical researchers and calibrated by three senior radiologists.
  • Focus: Specifically targets the L2-L3 vertebral levels, where fat distribution is most representative of overall health risks.
  • Accessibility: Publicly available on Figshare for non-commercial use.

Key Insights and Ideas

  • The "Intestinal Cavity" Confounder: One of the biggest hurdles in segmenting visceral fat is that the intestinal cavity often shares similar pixel characteristics (low brightness) with fat. This makes "simple" thresholding fail and necessitates the spatial awareness of deep learning.
  • BiSeNet as a "Dark Horse": While U-Net remains the gold standard for accuracy, the authors found that BiSeNet offered a compelling trade-off. It achieved segmentation results nearly as good as U-Net but with significantly shorter training times (approx. 2.5 hours vs. 10+ hours), making it ideal for real-time clinical applications.
  • Preservation vs. Smoothing: In denoising, they noted a critical tension: smoothing filters (like Mean/Median) produce better "numbers" (PSNR/SSIM) but often destroy the thin outer boundaries of subcutaneous fat. BM3D and Wavelet methods were superior at preserving the actual structural integrity of the medical image.
  • Population Phenotypes: The radiomics analysis wasn't just a math exercise; it successfully clustered the subjects into three distinct adipose distribution patterns, suggesting that this dataset can be used to identify specific patient risk groups.

Key Results

  • Segmentation Accuracy: U-Net led the pack with a Dice coefficient of 0.980 for subcutaneous fat and 0.949 for visceral fat.
  • Efficiency: BiSeNet achieved a Dice score of 0.978 (SAT) and 0.945 (VAT) while being the fastest model to train.
  • Denoising: For Poisson noise (common in CT), the Wiener filter and NLM algorithm performed best, achieving SSIM scores above 0.96.
  • Clustering: Unsupervised models (DBI index of 0.651 for K-Means) confirmed that the extracted radiomic features are robust enough to differentiate between different body compositions without human labels.

Limitations and Future Work

The authors acknowledge that while 300 subjects is a significant step forward for this niche, it is still small compared to general computer vision datasets. The current study also focuses primarily on 2D slices at the L2-L3 level; future work could expand into full 3D volumetric segmentation. Additionally, while they explored several radiomic features, the authors suggest that incorporating a wider array of "higher-order" features could further unlock the predictive power of the dataset for clinical diagnosis.

Summary updated

BibTeX
@article{ma2024aatct,
  author = {Zhiyu Ma and Chen Li and Tianming Du and Le Zhang and Dechao Tang and Deguo Ma and Shanchuan Huang and Yan Liu and Yihao Sun and Zhihao Chen and Jin Yuan and Qianqing Nie and Marcin Grzegorzek and Hongzan Sun},
  title = {{AATCT-IDS: A benchmark Abdominal Adipose Tissue CT Image Dataset for image denoising, semantic segmentation, and radiomics evaluation}},
  journal = {Computers in Biology and Medicine},
  year = {2024},
  doi = {10.1016/j.compbiomed.2024.108628},
}
Figure