Abstract
The assessment of abdominal adipose tissue distribution is critical for managing metabolic syndrome and cardiovascular risks. However, the development of automated diagnostic tools has been hindered by a lack of large-scale, high-quality annotated datasets. This paper introduces AATCT-IDS, a comprehensive benchmark dataset comprising 13,732 abdominal CT slices from 300 subjects. With expert-calibrated annotations for subcutaneous and visceral adipose tissue, AATCT-IDS provides a standardized platform for evaluating image denoising, semantic segmentation, and radiomics.
Introduction: Beyond BMI
Metabolic Syndrome (MetS) and obesity represent escalating global health crises. While clinical metrics such as Body Mass Index (BMI) and waist circumference are widely used, they are often insufficient for precise risk assessment. These "blunt instruments" fail to distinguish between subcutaneous adipose tissue (SAT)—fat stored just beneath the skin—and visceral adipose tissue (VAT)—fat surrounding internal organs. VAT is a significantly higher risk factor for metabolic and cardiovascular diseases.
Computed Tomography (CT) is the gold standard for visualizing these fat distributions. However, the field has long faced a "data desert." Existing datasets are frequently small, private, or lack the rigorous manual annotations required to train robust deep learning models. AATCT-IDS addresses this gap by providing a large-scale, open-source playground for researchers.
The AATCT-IDS Dataset
Sourced from Shengjing Hospital of China Medical University, the AATCT-IDS dataset consists of 13,732 raw CT slices from 300 subjects. To ensure high quality, 3,213 slices focusing on the L2-L3 vertebral levels—where fat distribution is most representative of overall health—were manually annotated.
The annotation process involved seven biomedical researchers and was calibrated by three senior radiologists. This rigorous approach ensures that the ground truth labels for SAT and VAT are clinically accurate, facilitating the development of high-precision segmentation models.
Methodology and Baseline Evaluations
The authors conducted an exhaustive baseline evaluation across three primary domains to demonstrate the dataset's utility:
1. Image Denoising
Medical CT images often suffer from various types of noise (Poisson, Gaussian, Speckle). The researchers tested seven classical and transform-domain algorithms, including BM3D, Wavelet Transforms, and Non-local Means (NLM). The goal was to identify which methods best preserve the delicate boundaries of adipose tissue while reducing noise.
2. Semantic Segmentation
Eight deep learning architectures were benchmarked, ranging from the foundational U-Net and FCN8s to modern, lightweight models like BiSeNet and ICNet. These models were tasked with automatically mapping SAT and VAT regions, a process complicated by anatomical variations.
3. Radiomics and Clustering
Beyond simple segmentation, the researchers extracted multidimensional features such as texture, area, and perimeter. Unsupervised learning techniques (K-Means, DBSCAN, and AGNES) were applied to determine if the data could naturally reveal distinct "phenotypes" of fat distribution within the population.
Key Challenges and Insights
A significant hurdle in visceral fat segmentation is the intestinal cavity. As shown in Figure 2, the intestinal cavity often shares similar low-brightness pixel characteristics with fat, causing simple thresholding methods to fail. This necessitates the use of deep learning models with strong spatial awareness.
Other key insights include:
- BiSeNet as an Efficient Alternative: While U-Net achieved the highest accuracy (Dice coefficient of 0.980 for SAT and 0.949 for VAT), BiSeNet emerged as a "dark horse." It provided nearly identical results (0.978 SAT, 0.945 VAT) with significantly reduced training times—approximately 2.5 hours compared to over 10 hours for U-Net—making it highly suitable for real-time clinical use.
- Structural Preservation in Denoising: Traditional smoothing filters like Mean or Median filters often destroy the thin outer boundaries of subcutaneous fat. In contrast, BM3D and Wavelet methods were superior at maintaining structural integrity.
- Identifying Risk Phenotypes: Radiomics analysis successfully clustered subjects into three distinct adipose distribution patterns, suggesting the dataset's potential for identifying specific patient risk groups through unsupervised learning.
Conclusion and Future Directions
AATCT-IDS represents a significant step forward in medical imaging for obesity research. By providing a large-scale, annotated benchmark, it enables the development of more accurate and efficient diagnostic tools. While the current dataset focuses on 2D slices at the L2-L3 level, future work aims to expand into full 3D volumetric segmentation and incorporate higher-order radiomic features to further enhance clinical predictive power.
Summary updated