🤖 AI

A Neighborhood Attention Transformer Network for Enhanced 3D Segmentation of the Left Anterior Descending Artery

This paper introduces NA-UNETR, a transformer-based segmentation framework utilizing Neighborhood Attention and uncertainty-guided optimization to significantly improve the accuracy of Left Anterior Descending artery delineation in low-contrast, free-breathing 3D CT scans for cardiac radiotherapy planning.

Original authors: Rafi Ibn Sultan, Chengyin Li, Yiannos Demetriou, Ahmed I. Ghanem, Joshua P. Kim, Justine Cunningham, Hassan Bagher-Ebadian, Dongxiao Zhu, Kundan S. Thind

Published 2026-08-13

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Rafi Ibn Sultan, Chengyin Li, Yiannos Demetriou, Ahmed I. Ghanem, Joshua P. Kim, Justine Cunningham, Hassan Bagher-Ebadian, Dongxiao Zhu, Kundan S. Thind

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a single, incredibly thin thread of gold hidden inside a giant, foggy block of jelly. That is roughly what doctors face when they try to map the "Left Anterior Descending" (LAD) artery on a standard CT scan of a patient's chest. This artery is a tiny, winding pipeline that feeds blood to the heart, but on a regular CT scan (which doesn't use special dye to make blood vessels pop), it looks almost invisible, blending right into the surrounding muscle. This is a huge problem for cancer patients getting radiation therapy. Doctors need to know exactly where this tiny artery is so they can aim the radiation beam to kill the tumor without accidentally frying the heart's fuel line. If they miss it, the patient could suffer heart trouble years later.

To solve this, scientists use "artificial intelligence" (AI) to act like a super-powered detective. Usually, these AI detectives are trained on millions of clear, high-contrast pictures. But here, the detectives only have a few blurry, foggy pictures to work with, and the "gold thread" is so small it's easy to miss. The paper you are about to read introduces a new type of AI detective called NA-UNETR. It's built on a clever idea called "Neighborhood Attention." Think of it like a detective who doesn't just stare at one spot in the fog but looks at a small neighborhood of pixels around it to figure out if a thread is passing through, while also remembering the big picture of the whole heart. The researchers wanted to see if this new detective could find the invisible artery better than the old, standard detectives, even when the clues were scarce and the picture was fuzzy.

The New Detective: NA-UNETR

The researchers, led by Rafi Ibn Sultan and his team, built a new AI model named NA-UNETR to tackle this "needle in a haystack" problem. Their goal was simple but difficult: make the AI see the LAD artery clearly in low-contrast CT scans where even human experts often disagree on where the artery actually is.

The secret sauce of NA-UNETR is its ability to look at the image in two ways at once. First, it uses Neighborhood Attention (NA). Imagine you are trying to trace a faint line in the fog. If you only look at one tiny dot, you might think it's just noise. But if you look at the little "neighborhood" of dots right next to it, you can see the line continuing. This part of the AI focuses on small, local details to keep the thin artery connected. Second, it uses Dilated Neighborhood Attention (DiNA). This is like taking a step back to see a wider neighborhood, allowing the AI to understand how the artery curves over a longer distance without losing the fine details. By combining these two "sights," the model captures both the tiny, fragile structure of the vessel and its long, winding path.

Training the Detective: The "Pre-Training" Trick

One of the biggest hurdles the team faced was a lack of data. They only had 20 patients with expertly drawn maps of the LAD artery to teach their AI. That's like trying to teach a student to drive with only 20 minutes of practice time. To fix this, they used a clever two-step training strategy.

First, they let the AI "pre-train" on a massive library of 1,000 clear, high-quality heart scans (called CTA) where the arteries are bright and easy to see. This taught the AI the general shape and structure of coronary arteries. Then, they took this knowledgeable AI and gave it a "fine-tuning" session using the 20 difficult, foggy scans. They didn't retrain the whole brain; instead, they used a technique called LoRA (Low-Rank Adaptation), which is like giving the AI a small, specialized notepad to learn the specific quirks of the foggy scans without forgetting everything it learned from the clear ones. This allowed them to get great results even with very little data.

The Results: Finding the Invisible Thread

When they put NA-UNETR to the test, the results were promising. On the difficult, foggy scans (the LAD-SEG dataset), the new model achieved a Dice score of 45.64%. To put that in perspective, the previous best models (like nnU-Net and Swin UNETR) only scored around 42% to 44%. While 45% might not sound like a perfect score, in the world of finding invisible arteries, it's a significant jump. The authors note that this improvement is comparable to the level of disagreement between different human doctors, suggesting the AI is now as good as a human expert at this tricky task.

The model also did a better job at finding the exact edges of the artery. It reduced the "Hausdorff Distance" (a measure of how far off the edge is) by about 2.96 mm compared to the Swin UNETR model. In the world of tiny arteries, being off by 3 millimeters is a huge difference between hitting the target and missing it.

When they tested the model on the clearer, public dataset (ImageCAS), it performed even better, hitting a Dice score of 79.49%, beating all other models including the heavy-hitters like UNet++ and Swin UNETR. This showed that the model wasn't just lucky; it actually learned how to see arteries well.

What the Model Did Not Do

It is important to note what the paper does not claim. The authors are very careful to say that while the AI is better, it is not perfect. They explicitly state that the task is still "fundamentally limited by the visibility of the anatomy." In other words, if the artery is too faint or the image is too blurry, even the best AI can't magically see it. The paper also notes that on the difficult foggy scans, the differences between the new model and the old ones were not statistically "proven" to be huge (due to the small number of patients), but the trend was clear. They do not claim this is ready for immediate use in every hospital tomorrow; they suggest it needs more testing and larger datasets before it can be trusted for routine clinical use.

The Bottom Line

The paper concludes that NA-UNETR is a robust and efficient way to find the Left Anterior Descending artery in difficult CT scans. By using a "neighborhood" approach to look at the image and a smart two-step training method, the model balances local detail with global context better than previous methods. It suggests that with this tool, doctors might eventually be able to spare the heart from radiation damage more effectively, though the authors remind us that the journey from a good computer model to a life-saving hospital tool still has some steps to go. The code for this new detective is now available for others to try and improve upon, keeping the search for the invisible thread going.

Technical Summary: NA-UNETR for LAD Segmentation

Problem Statement
Accurate segmentation of the Left Anterior Descending (LAD) artery in 3D free-breathing, non-contrast CT scans is critical for cardiac dose sparing in thoracic radiotherapy. However, this task is exceptionally difficult due to the LAD's minute caliber, poor soft-tissue contrast, and significant anatomical variability. Even manual delineation by experts suffers from low inter-observer agreement (Dice scores ranging from 0.10 to 0.53), highlighting the ambiguity of vessel boundaries. Existing automated methods, including standard CNNs (e.g., U-Net, nnU-Net) and Transformer-based models, struggle to capture the long-range anatomical dependencies required for vessel continuity while maintaining sensitivity to thin, low-contrast structures on non-contrast CT. Furthermore, the scarcity of expert-labeled LAD data creates a severe "small-data" bottleneck for supervised learning.

Methodology: NA-UNETR
The authors propose NA-UNETR, a 3D transformer-based segmentation framework designed to balance local precision with global context modeling. The methodology comprises four key technical components:

Architecture (Neighborhood Attention):
- The model utilizes a UNETR-style encoder-decoder backbone but replaces standard global self-attention with Neighborhood Attention (NA) and Dilated Neighborhood Attention (DiNA) blocks.
- NA restricts token interactions to local $k \times k \times k$ windows, introducing a spatial inductive bias that prevents the dilution of weak anatomical signals by dominant background responses. This is crucial for preserving the continuity of thin vascular structures.
- DiNA extends the receptive field progressively across layers by sampling dilated neighborhoods, allowing the model to integrate longer vessel segments without resorting to computationally expensive global interactions.
- The architecture includes residual convolution layers before each NAT block to stabilize training and depthwise convolutions to preserve fine-grained details.
Training Strategy (Two-Stage Transfer Learning):
- To address the scarcity of annotated LAD data, the model employs a two-stage approach. First, it is pretrained on a large public dataset of 1,000 contrast-enhanced CTA volumes (ImageCAS) to learn generalized coronary artery representations.
- Second, it is fine-tuned on the small institutional dataset (LAD-SEG, $N=20$ ) using Low-Rank Adaptation (LoRA). Specifically, the encoder weights are frozen, and only the decoder parameters and LoRA adapters (rank $r=8$ ) inserted into the MLP layers of the encoder are updated. This parameter-efficient strategy prevents overfitting on the small target dataset.
Loss Function (Uncertainty-Guided Optimization):
- A composite loss function combines Dice-Focal Loss (for region overlap and hard-example focusing) and Hausdorff Loss (for boundary precision).
- These components are dynamically balanced using homoscedastic uncertainty weighting. The model learns two log-variance parameters to automatically adjust the relative contribution of the segmentation and boundary losses during training, eliminating the need for manual weight tuning. A noise-perturbed term is added to the Hausdorff loss to encourage exploration of boundary features and prevent overfitting.
Data Processing:
- Preprocessing: Includes intensity clipping to $[-200, 400]$ HU, contrast enhancement via gamma adjustment, and artery-centric patch sampling (1:1 foreground-to-background ratio) to mitigate class imbalance.
- Postprocessing: A three-step pipeline retains the largest connected component, removes small objects (<64 voxels), and fills holes to ensure anatomical plausibility.

Key Contributions

Novel Architecture: Introduction of NA-UNETR, which integrates Neighborhood Attention into a 3D Transformer backbone to specifically address the challenges of thin, low-contrast tubular structures.
Small-Data Solution: A principled transfer learning framework combining large-scale CTA pretraining with LoRA-based fine-tuning, effectively overcoming the bottleneck of limited annotated LAD data.
Optimization Strategy: The use of homoscedastic uncertainty to dynamically balance segmentation and boundary losses, improving both volumetric overlap and contour accuracy without manual hyperparameter tuning.

Results
The model was evaluated on the institutional LAD-SEG dataset and the public ImageCAS dataset against state-of-the-art CNN and Transformer baselines (including nnU-Net, Swin UNETR, and UNETR).

On LAD-SEG (Non-contrast CT): NA-UNETR achieved a Dice Similarity Coefficient (DSC) of 45.64%, improving by 3.10 percentage points over nnU-Net (42.54%). It also achieved the best boundary metrics among all models, with an HD95 of 38.16 mm and ASD of 10.01 mm. While the absolute Dice score remains modest (consistent with the difficulty of the task and inter-observer variability), NA-UNETR demonstrated superior boundary accuracy and geometric stability compared to competitors.
On ImageCAS (Contrast CT): The model achieved a DSC of 79.49%, outperforming all baselines (e.g., UNet++ at 78.57%) and showing statistically significant improvements ( $p < 0.05$ ) in Dice, clDice, HD95, and ASD.
Efficiency: NA-UNETR maintained computational efficiency comparable to Swin UNETR (19.6M parameters, 314.1B FLOPs) with moderate GPU memory usage (4.17 GB) and inference latency (1.33 s/volume).
Ablation Studies: Confirmed that the combination of residual blocks, variable kernel sizes, the two-stage training strategy, and the uncertainty-weighted loss function were all critical to performance stability and boundary precision.

Significance and Claims
The paper claims that NA-UNETR effectively balances local precision and global context, enabling more accurate segmentation of the challenging LAD structure in non-contrast CT than previous methods. The authors position this work as a robust and computationally efficient framework for substructure-level cardiac segmentation in radiotherapy planning.

The authors are modest in their clinical claims, acknowledging that:

The achieved performance on non-contrast CT is still bounded by the intrinsic visibility limits of the anatomy in these images.
While the model improves upon baselines and reduces the gap with expert variability, it is not yet suitable for routine clinical deployment.
The primary value lies in reducing manual editing effort and improving consistency for small, elongated vascular structures, laying the groundwork for future integration into cardiac dose assessment workflows.
Future work requires larger annotated datasets, multi-institutional validation, and prospective studies to evaluate time savings and user acceptance before clinical adoption.