ImageTBAD: 3D CTA Data for TBAD Segmentation

• ImageTBAD is a 3D CTA dataset curated for TBAD, offering manual annotations for true lumen (TL), false lumen (FL), and the challenging false lumen thrombus (FLT).
• The dataset incorporates standardized preprocessing, augmentation, and three-fold cross-validation to support robust deep learning benchmarking.
• Quantitative evaluations reveal competitive Dice scores for TL and FL, while FLT segmentation remains challenging, highlighting avenues for future research.

ImageTBAD is a publicly released dataset comprising 3D computed tomography angiography (CTA) images specifically curated for the automatic segmentation of Type-B Aortic Dissection (TBAD). It is the first resource to include manual annotations for not only true lumen (TL) and false lumen (@@@@3@@@@), but also false lumen thrombus (@@@@1@@@@)—a clinically critical yet technically challenging segmentation target. The dataset, developed by researchers at Guangdong Provincial People’s Hospital, aims to facilitate and benchmark advances in deep learning–based techniques for the quantitative assessment of TBAD, with particular emphasis on the automated delineation of variable and irregular FLT regions (Yao et al., 2021).

1. Dataset Composition and Preprocessing

ImageTBAD contains 100 pre-operative thoracic aortic CTA volumes, collected prospectively from January 2013 to April 2015 using Siemens SOMATOM Force and Philips 256-slice Brilliance iCT scanners. The cohort includes 31% female patients, with a mean age of $52.5 \pm 11.3$ years. Each scan provides a $512 \times 512$ in-plane grid, with 135–416 slices per volume and a voxel spacing of $0.25 \times 0.25 \times 0.25$ mm³; slice spacing is 0.75 mm. Contrast-enhanced imaging spans from the aortic root to the supra-aortic vessels.

Standardized preprocessing includes:

• Intensity normalization (zero-mean, unit-variance scaling)
• Bounding-box proposal resampling to $64^3$ voxels
• Final region-of-interest (ROI) cropping and resampling to $S \times S \times 2S$ ($S=64$ or $96$)
• Image augmentation (random rotation, scaling, elastic deformation) as per Payer et al. 2017

This systematic approach supports robust model training and facilitates cross-study benchmarking.

2. Annotation Protocol and Label Definitions

Three anatomical labels are systematically annotated:

• True lumen (TL): Patent channel distal to the dissection tear.
• False lumen (FL): Blood-filled aortic channel between intima and media, excluding thrombus.
• False lumen thrombus (FLT): Low-attenuation clot within part of the FL.

Labeling protocol entails manual segmentation by a senior cardiovascular radiologist (1–1.5 hours/scan), with review and consensus from a second expert. Inter-observer agreement is described as high after consensus review, though no explicit kappa statistic is reported.

FLT segmentation introduces distinct challenges:

• High inter-patient morphological variability: FLT may occur at diverse positions (top, middle, bottom, diffuse along the aorta, or as separated patches).
• Poor tissue contrast: FLT often presents minimal intensity difference relative to adjacent soft tissue or FL, leading to ambiguous boundaries requiring expert judgment.

3. Data Partitioning and Distribution

The dataset is partitioned into three folds for cross-validation, each containing approximately 67 training and 33 test cases. Notably, the ratio of FLT-positives is balanced between folds (68 with FLT, 32 without). No dedicated held-out validation set is used; all reported metrics average across the three cross-validation splits.

ImageTBAD, along with baseline segmentation code, is available for non-commercial research as DICOM/NIfTI files. Each subject directory contains:

• image.nii.gz: Raw CTA volume
• label.nii.gz: Segmentation labels (0=background, 1=TL, 2=FL, 3=FLT)

4. Baseline Segmentation Framework

The authors introduce a two-stage segmentation pipeline based on 3D U-Net architectures:

• ROI Extraction: Full CTA volumes are downsampled to $64^3$ and segmented with a four-level 3D U-Net (feature maps: $N$, $2N$, $4N$, $8N$) to localize the aggregated aorta (TL+FL+FLT). A bounding box is then extracted and resampled for refined analysis.
• ROI Segmentation: Two approaches are explored:
  • Approach A (Multi-task): Refinement is split into (i) aorta mask refinement, followed by (ii) joint segmentation of TL and FL. FLT is obtained by set subtraction: $$\mathrm{FLT} = \mathrm{Aorta} - (\mathrm{TL} \cup \mathrm{FL})$$.
  • Approach B (Simultaneous): Direct segmentation of TL, FL, and FLT as separate classes.

Network optimization employs a weighted loss function: $$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{Dice}} + \mathcal{L}_{\text{CE}}$$ with

$$\mathcal{L}_{\text{Dice}} = 1 - \frac{2\sum_i p_i\,g_i}{\sum_i p_i + \sum_i g_i}$$

$\mathcal{L}_{\text{CE}} = -\sum_i g_i \log(p_i)$

where $p_i$ is the softmax probability for voxel $i$, $g_i \in \{0,1\}$ is the ground-truth indicator. Training uses Adam (lr=$10^{-4}$), five epochs per subnetwork, batch sizes of 4 (S=64, N=64) or 3 (S=96, N=32). Data augmentation is as above.

5. Quantitative Evaluation and Comparative Analysis

Performance is evaluated using:

• Dice coefficient for each anatomical class:

$$\mathrm{Dice}(G,S) = \frac{2\,|G \cap S|}{|G| + |S|}$$

• Hausdorff distance (HD) for boundary errors:

$$\mathrm{HD}(G,S) = \max\left\{\sup_{x\in\partial G}\inf_{y\in\partial S}||x-y||,\,\sup_{y\in\partial S}\inf_{x\in\partial G}||x-y||\right\}$$

Key results for Approach B ($S=96$), averaged over three folds:

• Aorta: Dice = $0.91 \pm 0.04$, HD = $667.4 \pm 612.3$
• TL: Dice = $0.85 \pm 0.07$, HD = $288.4 \pm 426.0$
• FL: Dice = $0.78 \pm 0.21$, HD = $643.7 \pm 1999.5$
• FLT: Dice = $0.52 \pm 0.40$, HD = $978.6 \pm 2887.3$

For cases without FLT (32 volumes), perfect FLT Dice ($1.0$) is achieved by not predicting FLT. However, in the 68 positive cases, FLT Dice drops to ~0.20 on average, meaning overall FLT segmentation remains challenging.

Comparison with prior approaches:

• Li et al. (3D multi-task CNN): Dice (aorta/TL/FL) = $0.910/0.849/0.821$
• Cao et al. (FCN): Dice (aorta/TL/FL) = $0.93/0.93/0.91$

Aorta segmentation is competitive; TL and FL are slightly lower than in dedicated methods; and FLT remains a systematically under-addressed challenge, due to its small volume, irregular distributions, low contrast, and weak anatomical priors.

6. Distribution and Recommendations for Future Work

ImageTBAD and baseline code are distributed publicly for academic use (see “our dataset” link in (Yao et al., 2021)). Provided resources include documented instructions for environment setup (PyTorch $\geq$1.0), preprocessing, and training.

Enhancements recommended by the authors include:

• Attention-gated or multi-scale networks (e.g., nnU-Net with attention modules) to focus on subtle FLT regions.
• Boundary-aware or topology-preserving loss functions.
• Cascaded detection–segmentation schemes with specialized FLT candidate proposal mechanisms.
• Integration of centerline extraction and shape priors to regularize FLT geometry.
• Augmentation with a larger dataset and multi-phase CT for improved thrombus-lumen contrast.

7. Significance and Research Context

ImageTBAD is the first publicly available 3D CTA dataset to provide comprehensive annotations of TL, FL, and FLT. Its design exposes the persistent segmentation gap in FLT detection—reflected in a modest mean Dice of $\sim 0.52$ overall and $\sim 0.20$ on positive cases—thereby motivating new methodological research into region- and boundary-sensitive learning frameworks for clinically relevant TBAD imaging tasks (Yao et al., 2021). The dataset represents a benchmark for automated aortic dissection analysis and presents a uniquely challenging resource for medical image computing.