BiM-GeoAttn-Net: Linear-Time Depth Modeling with Geometry-Aware Attention for 3D Aortic Dissection CTA Segmentation

The paper proposes BiM-GeoAttn-Net, a lightweight framework that combines Bidirectional Depth Mamba for efficient cross-slice modeling and a Geometry-Aware Vessel Attention module to achieve robust, high-accuracy 3D segmentation of aortic dissection lumens in CTA scans.

The Gist

Imagine you are trying to trace a winding, slippery river through a thick, foggy forest using a series of 2D aerial photographs. This is essentially what doctors face when they try to map an Aortic Dissection (a tear in the body's main artery) using CT scans.

The artery is like that river: it twists, turns, and sometimes looks very faint against the surrounding tissue. The CT scans are like a stack of 2D photos. If you look at just one photo, the river might look broken or blurry. If you look at the whole stack, you need to make sure the river connects smoothly from the top photo to the bottom one.

Here is how the paper's new invention, BiM-GeoAttn-Net, solves this problem, explained simply:

The Problem: The "Foggy River" Challenge

Current computer programs (AI) that try to draw this artery often make two big mistakes:

1. The "Broken River" Effect: Because the photos are taken slice-by-slice, the AI often loses track of the river between slices. One slice shows the river, the next shows a gap, and the next shows it again. The result is a jagged, disconnected mess instead of a smooth tube.
2. The "Faint Outline" Problem: In some areas, the river looks very similar to the surrounding mud (tissue). The AI gets confused and either draws the river too wide, too narrow, or misses it entirely.

The Solution: A Smart Two-Step Team

The authors created a new AI team called BiM-GeoAttn-Net. Think of it as a two-person inspection crew working together to fix the map.

Step 1: The "Stack Reader" (Bidirectional Depth Mamba)

• The Metaphor: Imagine you are reading a book, but instead of reading left-to-right, you are reading a stack of pages from top to bottom and bottom to top at the same time.
• What it does: This part of the AI looks at the entire stack of CT slices together. It uses a special, fast math trick (called "Mamba") to understand how the artery in one slice connects to the slice above and below it.
• The Result: It fixes the "Broken River" problem. Even if one slice is blurry, this module knows, "Hey, the river was here in the slice above and here in the slice below, so it must be here too." It ensures the artery looks like one long, continuous tube, not a string of disconnected beads.

Step 2: The "Detail Sharper" (Geometry-Aware Vessel Attention)

• The Metaphor: Imagine you are trying to trace a thin wire on a table, but the table is covered in dust. A normal flashlight just lights up everything equally. This module is like a specialized laser pointer that only shines on things that look like long, thin tubes.
• What it does: Once the "Stack Reader" has connected the dots, this module zooms in on the edges. It knows that arteries are long, round tubes. It ignores the "dust" (background noise) and sharpens the edges of the "wire" (the artery). It acts like a high-precision editor that cleans up fuzzy lines.
• The Result: It fixes the "Faint Outline" problem. The edges of the artery become crisp and clear, even in the foggiest parts of the scan.

Why is this a big deal?

Most previous AI models were either:

• Too slow: Like trying to read the whole book by looking at every single letter individually (Transformers).
• Too short-sighted: Like only looking at one page at a time (standard CNNs).

This new model is fast and smart. It reads the whole stack quickly (linear time) and then sharpens the details.

The Results

When the researchers tested this new tool on real patient data:

• It was more accurate than the current best methods at finding the exact shape of the artery.
• It made fewer mistakes where the artery looked broken or disconnected.
• It did all this without needing a supercomputer; it runs efficiently on standard medical equipment.

In a Nutshell

BiM-GeoAttn-Net is like giving a doctor a pair of smart glasses. One lens helps them see the whole river from start to finish without losing the path, and the other lens helps them see the exact edges of the riverbank, even in the fog. This helps doctors make better decisions about how to treat patients with dangerous aortic tears.

Technical Summary

1. Problem Statement

Accurate segmentation of aortic dissection (AD) lumens in CT Angiography (CTA) is critical for clinical decision-making and morphological assessment. However, existing methods face two primary challenges:

• Inter-slice Incoherence: Due to anisotropic voxel spacing and the complex, tortuous geometry of the aorta, standard 3D CNNs often fail to maintain continuity across slices, leading to fragmented lumen predictions.
• Low-Contrast Boundary Ambiguity: The boundaries between the dissected lumen and surrounding tissues are often subtle. Existing models struggle to refine these boundaries without introducing noise or losing fine structural details.
• Computational Limitations: While Transformers offer global context, full 3D self-attention is computationally prohibitive. Conversely, standard State-Space Models (SSMs) often lack explicit alignment with anatomical continuity (e.g., the depth axis) and structural priors necessary for low-contrast refinement.

2. Methodology: BiM-GeoAttn-Net

The authors propose BiM-GeoAttn-Net, a lightweight framework built upon the 3D nnU-Net architecture. It enhances the bottleneck of the encoder-decoder structure with two novel, cascaded modules designed to address the specific challenges of AD segmentation.

A. Bidirectional Depth Mamba (BiM)

• Objective: To model long-range dependencies along the depth (slice) axis efficiently, ensuring inter-slice coherence.
• Mechanism:
  • Instead of applying expensive 3D self-attention, the input feature map is reshaped into depth-wise sequences ($D$-axis).
  • A Mamba state-space model (SSM) is applied bidirectionally (forward and backward) to capture cross-slice context.
  • The forward ($z_{\rightarrow}$) and backward ($z_{\leftarrow}$) states are fused and added residually to the input.
  • A lightweight MLP with pre-normalization further refines the features.
• Advantage: Achieves linear-time complexity with respect to depth ($D$), significantly reducing computational cost compared to quadratic 3D attention while stabilizing slice-to-slice continuity.

B. Geometry-Aware Vessel Attention (GeoAttn)

• Objective: To refine tubular structures and sharpen ambiguous boundaries using orientation-sensitive priors.
• Mechanism:
  • Anisotropic Filtering: The module employs three plane-aligned convolutions ($1\times3\times3$, $3\times1\times3$, $3\times3\times1$) alongside a standard $3\times3\times3$ convolution. This captures orientation-specific features of tubular vessels.
  • Dual Attention:
    1. Spatial Attention: A lightweight bottleneck generates a spatial attention map modulated by a residual gate to emphasize vessel regions.
    2. Channel Attention: Utilizes global average and max pooling to weight feature channels effectively.
• Advantage: By combining direction-aware filtering with attention mechanisms, GeoAttn suppresses background interference and sharpens boundaries that are often lost in low-contrast regions.

C. Training Strategy

• Loss Function: A hybrid objective combining Dice Loss and Cross-Entropy (CE) to address foreground-background imbalance.
• Architecture: The enhanced features from the BiM-GeoAttn bottleneck are residually fused and passed to the decoder for final segmentation.

3. Key Contributions

1. Novel Framework: Introduction of BiM-GeoAttn-Net, a specialized 3D segmentation framework tailored for the specific anatomical and imaging challenges of aortic dissection.
2. Bidirectional Depth Modeling: Design of the BiM module, which leverages linear-time SSMs to enforce cross-slice consistency without the computational overhead of full 3D attention.
3. Geometry-Aware Refinement: Development of the GeoAttn module, which utilizes orientation-sensitive anisotropic filtering and dual attention to enhance tubular structure delineation and boundary precision.
4. Comprehensive Validation: Extensive experiments demonstrating superior performance over CNN, Transformer, and SSM-based baselines in both volumetric overlap and boundary accuracy.

4. Experimental Results

The method was evaluated on a multi-source dataset (Dataset500_VesselTBAD) containing 71 Type-B Aortic Dissection cases (split 50/7/14 for train/val/test).

• Quantitative Performance:
  • Dice Score: Achieved 93.35%, outperforming the best baseline (nnU-Net at 90.84%) by 2.51%.
  • Boundary Accuracy (HD95): Achieved 12.36 mm, a significant improvement over nnU-Net (18.51 mm) and SegFormer3D (24.15 mm), though slightly higher than Attention U-Net (10.77 mm).
  • Efficiency: Training time was 2.9 min/epoch with 8.0 GB memory usage, comparable to nnU-Net and significantly more efficient than Transformer-based models (e.g., SegFormer3D at 4.6 min/epoch).
• Ablation Studies:
  • Adding BiM alone improved Dice to 91.12% and reduced HD95 to 20.35 mm, confirming its role in fixing slice discontinuity.
  • Adding GeoAttn alone improved local structure sensitivity.
  • The combined model yielded the best results, proving the complementarity of global depth modeling and local geometric refinement.
• Qualitative Results: Visual comparisons showed that BiM-GeoAttn-Net produced fewer false positives/negatives, maintained better lumen continuity, and offered sharper boundaries compared to CNN, Transformer, and standard SSM baselines.

5. Significance and Impact

• Clinical Relevance: The proposed method provides a robust, automated solution for segmenting complex aortic dissections, which is essential for quantitative morphological analysis and longitudinal monitoring of patients.
• Efficiency vs. Accuracy: The work demonstrates that linear-time depth modeling coupled with geometry-aware attention can achieve state-of-the-art accuracy without the heavy computational burden of 3D Transformers.
• Generalizability: The approach of aligning sequence modeling direction with anatomical continuity (depth axis) and incorporating structural priors (vessel orientation) offers a promising paradigm for other 3D medical imaging tasks involving tubular structures or anisotropic data.