Intracranial aneurysm segmentation with nnU-net: utilizing loss functions and automated vessel extraction

Maysam Orouskhani, Negar Firoozeh, Huayu Wang, Yan Wang, Hanrui Shi, Weijing Li, Beibei Sun, Jianjian Zhang, Xiao Li, Huilin Zhao, Mahmud Mossa‐Basha, Jenq–Neng Hwang, Chengcheng ZhuVessel 2025 · DOI

TL;DR

Researchers have optimized the popular nnU-Net framework for intracranial aneurysm segmentation by integrating hybrid loss functions and automated vessel-aware cropping. By focusing the model's "attention" on vascular structures and prioritizing "hard" pixels during training, the team achieved significant performance gains, securing a 4th-place ranking in the international ADAM challenge.

What Problem Does This Paper Address?

Detecting and segmenting unruptured intracranial aneurysms (UIAs) is a "needle in a haystack" problem. These life-threatening bulges in brain arteries are often tiny (median ~3.6 mm) compared to the vast volume of a 3D Magnetic Resonance Angiography (MRA) scan.

Standard deep learning models struggle with two main issues here:

  1. Extreme Class Imbalance: Aneurysm voxels make up a microscopic fraction of the total image data.
  2. Structural Complexity: Aneurysms are physically attached to blood vessels and often mimic their intensity and shape, making it difficult for models to delineate where the vessel ends and the aneurysm begins.

While manual segmentation is the gold standard, it is labor-intensive and prone to human error. This paper seeks an automated, robust pipeline that can handle diverse aneurysm sizes and shapes across different clinical datasets.

Methods

The authors built their solution on top of nnU-Net (the "no-new-Net" framework), which is the current state-of-the-art for medical image segmentation. They introduced three key enhancements:

  • Hybrid Loss Functions: Instead of relying solely on standard Dice or Cross-Entropy loss, they experimented with combinations including TopK, Focal, and Tversky losses. The goal was to force the model to focus on "hard" voxels (those the model is most unsure about) and better handle the class imbalance.
  • Automated Vessel-Based Cropping: To reduce noise and computational load, the team developed an unsupervised algorithm using a Hessian matrix and Otsu’s thresholding to identify brain vessels. They then cropped the 3D images to a Region of Interest (ROI) containing only these vascular structures, effectively removing ~60% of irrelevant background tissue.
  • Multi-Modality Training: The model was trained using two input channels: the original TOF-MRA image and a "vessel-enhanced" version (derived from the Hessian filter). This provides the network with explicit structural cues about the vascular tree.

Datasets

The study utilized two distinct datasets to prove the model's versatility:

  1. ADAM (Aneurysm Detection And SegMentation): A public challenge dataset with 113 cases. These aneurysms are generally small (median diameter 3.6 mm).
  2. RENJI: A private dataset from Ren Ji Hospital (213 cases) featuring larger, more complex aneurysms (median diameter 9.35 mm), including ruptured cases and those with internal "dark signal" regions.

Key Insights and Ideas

  • The "TopK" Advantage: The most successful loss configuration was a Weighted Dice + TopK + CE combo. The "TopK" component is particularly clever—it calculates loss only on the most difficult pixels, preventing the model from becoming "lazy" by over-optimizing on easy-to-segment background areas.
  • Cropping as Attention: The authors found that automated cropping didn't just speed up the model; it improved accuracy. By physically removing non-vascular tissue, they essentially provided the model with "hard-coded attention," preventing it from misidentifying noise in other brain regions as an aneurysm.
  • Complexity Matters: The multi-modality approach (using vessel-enhanced images) showed much greater benefits on the RENJI dataset than on ADAM. This suggests that as clinical cases become more complex (larger, irregular shapes), providing explicit vascular context becomes increasingly critical.

Key Results

The optimized pipeline demonstrated strong performance across both benchmarks:

  • RENJI Dataset: Achieved a Dice coefficient of 0.72 and a Sensitivity of 0.69.
  • ADAM Dataset: Achieved a Dice coefficient of 0.54 and a Sensitivity of 0.53.
  • Challenge Performance: In the official ADAM challenge, the proposed method reached 4th place in the segmentation task, achieving a 39% Dice score on the hidden testing set (where the top-ranking entry was 43%).
  • Vessel Focus: The automated cropping reduced the search space significantly, leading to a more stable training process and tighter confidence intervals for the results.

Limitations and Future Work

Despite the success, the authors noted several areas for improvement:

  • Dataset Bias: The ADAM dataset is skewed toward very small aneurysms, which may limit the model's generalizability to larger, more varied clinical cases without further tuning.
  • Modality Constraints: The study focused exclusively on TOF-MRA. Future work should explore whether these techniques translate to other common imaging types like CT Angiography (CTA) or Contrast-Enhanced MRA.
  • Sample Size: While 326 cases is a respectable start, "big data" in medical imaging often requires thousands of cases to ensure the model can handle rare aneurysm morphologies and different scanner hardware.

Summary updated

Abstract

The proposed method offers a promising solution for accurately segmenting intracranial aneurysms, showcasing superior performance compared to existing approaches. By integrating hybrid loss functions and vessel information, we enhance the model's ability to delineate intricate aneurysm structures, contributing to improved diagnosis and treatment planning for patients with intracranial aneurysms.
BibTeX
@inproceedings{orouskhani2025intracranial,
  author = {Maysam Orouskhani and Negar Firoozeh and Huayu Wang and Yan Wang and Hanrui Shi and Weijing Li and Beibei Sun and Jianjian Zhang and Xiao Li and Huilin Zhao and Mahmud Mossa‐Basha and Jenq–Neng Hwang and Chengcheng Zhu},
  title = {{Intracranial aneurysm segmentation with nnU-net: utilizing loss functions and automated vessel extraction}},
  booktitle = {Vessel},
  year = {2025},
  doi = {10.20517/2574-1209.2025.42},
}
Figure