BTCV Dataset for Abdominal CT Segmentation

• BTCV dataset is a benchmark of 30 contrast-enhanced CT scans annotated at the voxel level for 13 abdominal organs, enabling detailed segmentation studies.
• It employs standardized preprocessing techniques such as resampling, intensity normalization, and spatial cropping to support both transformer and CNN segmentation models.
• State-of-the-art methods like UNETR and DeepEdit demonstrate its utility by achieving high Dice scores, advancing automated and interactive segmentation research.

The Multi-Atlas Labeling Beyond the Cranial Vault (BTCV) dataset is a benchmark collection of contrast-enhanced abdominal CT scans with voxel-wise multi-organ segmentations. It serves as a standard testbed for automated and interactive 3D medical image segmentation, particularly in the development of deep learning-based methodologies. The dataset is widely referenced in the literature for its rigorous multi-class labeling protocol, variety of anatomic structures, and relevance for clinical and algorithmic assessment in abdominal analysis (Diaz-Pinto et al., 2023, Hatamizadeh et al., 2021).

1. Dataset Composition and Annotation Protocol

The BTCV dataset comprises 30 anonymized abdominal CT volumes, each annotated at the voxel level with ground-truth segmentations for 13 distinct organs and vessels: spleen, right kidney, left kidney, gallbladder, esophagus, liver, stomach, aorta, inferior vena cava, portal & splenic veins, pancreas, adrenal glands, and the background class (implicit in most protocols). The imaging modality is contrast-enhanced CT, acquired in the portal venous phase, with each scan containing 80–225 axial slices of 512×512 pixels and slice thicknesses ranging from 1 to 6 mm. All volumes are resampled to isotropic voxels—typically 1.0 mm³ for standard transformer models and 1.5 mm³ for segmentation approaches focusing on computational efficiency (Hatamizadeh et al., 2021, Diaz-Pinto et al., 2023).

Label maps are one-hot encoded into 13 channels, and all structures receive expert segmentation, enabling rigorous per-organ and global evaluation.

2. Preprocessing and Data Augmentation

Preprocessing protocols commonly include:

• Resampling: Volumes standardized to 1.0×1.0×1.0 mm³ or 1.5×1.5×1.5 mm³ isotropic spacing depending on model requirements (Hatamizadeh et al., 2021, Diaz-Pinto et al., 2023).
• Intensity Normalization: Intensity values are typically clipped to [−1000, 1000] HU and linearly scaled to [0, 1] for transformer pipelines; for certain interactive segmentation protocols, intensity clipping to [−125, 275] HU and per-volume z-score normalization are used (Diaz-Pinto et al., 2023).
• Spatial Cropping: Random crops of fixed size (e.g., 128³ voxels) are extracted to focus on the abdominal cavity during training.
• Data Augmentation: Techniques include random flips, affine rotations (e.g., ±15° jitter or 90°, 180°, 270° rotations), intensity scaling/shifts, and additive Gaussian noise. These augmentations address the limited case count and anatomical variability (Hatamizadeh et al., 2021, Diaz-Pinto et al., 2023).

3. Benchmark Tasks and Evaluation Metrics

BTCV is primarily used as a testbed for multi-organ segmentation, including both fully automatic and (semi-)interactive approaches. Typical challenges include the delineation of small-volume structures (e.g., adrenal glands), organ shape variability, and inter-class proximity.

Evaluation metrics are standardized:

• Dice coefficient for per-organ and average segmentation accuracy.
• 95% Hausdorff Distance (HD95) for boundary accuracy (though not always reported in interactive segmentation studies).
• Reporting adheres to both single-organ (e.g., spleen) and multi-label (e.g., 4-organ or full 13-organ) configurations.

Qualitative comparison and model ablation studies further support quantitative evaluation in the literature (Hatamizadeh et al., 2021, Diaz-Pinto et al., 2023).

4. State-of-the-Art Segmentation Methods Utilizing BTCV

BTCV has underpinned the validation of several seminal architectures:

• UNETR: Employs a transformer encoder over 16³ 3D patches and a U-Net style convolutional decoder. Direct skip connections from multiple transformer layers facilitate multi-scale feature fusion, improving performance over both CNN-only and earlier hybrid methods. In the Standard BTCV competition (30 training cases), UNETR achieves an average Dice coefficient of 0.856, and in the Free competition (80 cases), 0.891—ranking among state-of-the-art (Hatamizadeh et al., 2021).
• DeepEdit: Integrates automatic segmentation (UNETR backbone) with interactive refinement through simulated/user clicks, yielding a flexible auto-to-interactive spectrum. DeepEdit incorporates uncertainty estimation (aleatoric and epistemic) and active learning ranking, operationalizing annotator-in-the-loop paradigms. On BTCV, DeepEdit recovers state-of-the-art Dice scores (spleen: up to 0.931 with 10 clicks in fully interactive mode), with auto mode lagging slightly behind task-specific UNETR baselines until interactive input is provided (Diaz-Pinto et al., 2023).

Other methods frequently referenced on the BTCV leaderboard include nnU-Net, TransUNet, and CoTr, with varying architectures and decoder designs (Hatamizadeh et al., 2021).

Table: Excerpt of Validation Dice Scores for Key Methods on BTCV

Scheme/Model	Spleen Dice	Multilabel Dice (4-organs)	Competition Setting
UNETR baseline	0.919±0.017	0.911±0.021	Standard
DeepEdit-0.25	0.902±0.019 (1 click)	0.900±0.018 (5 clicks)	Interactive
DeepEdit-0 (DeepGrow)	0.931±0.010 (10 clicks)	0.926±0.011 (10 clicks)	Fully interactive

5. Training Protocols and Best Practices

Training on BTCV standardly employs hybrid loss functions (soft Dice + voxel-wise cross-entropy), Adam-based optimizers (Adam or AdamW), and data augmentation schemes tailored to anatomical variability and dataset size.

For UNETR, recommended hyperparameters include a patch size of 16³, embedding dimension of 768, 12 transformer layers/heads, batch size of 6, and 20,000 training iterations. Learned 1D positional embeddings, hybrid loss, and sliding-window inference (50% overlap) are critical for optimal performance. Five-fold cross-validation and ensembling further enhance generalizability and leaderboard rankings (Hatamizadeh et al., 2021).

Interactive segmentation frameworks simulate user clicks as Gaussian heatmaps concatenated to the image channels, with click-based training and mixed auto/interactive iterations (parameterized by the probability $p$ of auto mode) (Diaz-Pinto et al., 2023).

6. Applications and Impact on Segmentation Research

BTCV is foundational in evaluating algorithms for:

• Automated abdominal organ segmentation for clinical workflow and downstream tasks.
• Interactive and active learning frameworks where annotation efficiency, annotator effort reduction, and uncertainty-driven volume ranking are critical (Diaz-Pinto et al., 2023).
• Transformer-based 3D segmentation and the benchmarking of architectural innovations that enable long-range spatial dependency modeling within volumetric medical image data (Hatamizadeh et al., 2021).

The dataset’s granularity and anatomical diversity facilitate robust comparisons of model design, preprocessing, and human-in-the-loop annotation strategies.

7. Strengths, Limitations, and Ongoing Directions

BTCV’s strengths lie in its high-quality, multi-organ expert annotations, standardized CT imaging protocol, and adoption as a segmentation benchmark for both conventional and data-efficient deep learning paradigms. It is especially valuable for assessing performance on challenging, small organs.

Limitations highlighted in recent studies include restricted case number (30 volumes), lack of time-motion or human annotation studies specific to BTCV, and sparse reporting of boundary-sensitive metrics like Hausdorff distance in interactive settings. Further user studies are needed to quantify annotation time savings and clinical workflow integration for modern interactive methods.

BTCV continues to shape best practices in preprocessing, data augmentation, model selection, and evaluation in the field of abdominal CT organ segmentation (Diaz-Pinto et al., 2023, Hatamizadeh et al., 2021).