💻 computer science

Lesion Segmentation in Moderate to Severe Traumatic Brain Injury: An nnU-Net Based Approach with Adaptive Normalization in the AIMS-TBI 2025 Challenge

This paper presents a deep learning solution for the AIMS-TBI 2025 Challenge that utilizes an nnU-Net framework enhanced with adaptive brain-parenchyma normalization to achieve competitive performance in segmenting heterogeneous lesions in moderate to severe traumatic brain injury.

Original authors: Inhwa Son, Gaeun Lee, Sohyeon Sim, Kwang-Hyun Uhm

Published 2026-07-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Inhwa Son, Gaeun Lee, Sohyeon Sim, Kwang-Hyun Uhm

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 your brain is a bustling city, and a traumatic brain injury (TBI) is like a sudden, chaotic storm that leaves behind a messy mix of potholes, flooded streets, and collapsed buildings. Some of these damages are huge and obvious, while others are tiny, scattered cracks hidden in the pavement. The problem? Every storm leaves a different mess. One city might have a giant crater, while another has a thousand tiny cracks. Trying to map these damages automatically is like trying to use a single, rigid ruler to measure a city that keeps changing its shape and size.

This is the exact puzzle the AIMS-TBI 2025 Challenge set out to solve. Researchers gathered MRI scans from 875 patients across 13 different countries, creating a massive, messy dataset where the "damage" (lesions) looked wildly different from person to person. The goal? To build a computer program that could automatically draw a line around these injuries on brain scans.

The "One-Size-Fits-All" Mistake

For a long time, computer programs tried to "normalize" (or standardize) these brain scans by looking at the entire image, including the skull and the empty space outside the head. Think of this like trying to adjust the volume on a radio by listening to the static from the empty air around the station, not just the music. Because every person's skull is a different thickness and brightness, this method confused the computer. It was like trying to hear a whisper in a room where the walls were constantly changing color.

The authors of this paper argued against this old way of doing things. They showed that when you include the skull and background in your calculations, you introduce too much "noise," making it harder for the AI to spot the actual injuries.

The "Brain-Only" Filter

Instead, the team built a clever new strategy called Adaptive Normalization. Imagine you are a detective trying to find clues in a room. Instead of looking at the whole house (including the attic and the basement), you put on a special pair of glasses that only lets you see the living room where the clues are.

In their method, the computer first draws a mask to isolate just the brain tissue (the "parenchyma"), ignoring the skull and everything else. It then calculates the brightness and contrast only based on that brain tissue. This is like adjusting the radio volume based only on the music, ignoring the static. By doing this, the computer could see the differences between healthy brain tissue and the injured spots much more clearly, even when the scans came from different hospitals with different machines.

The Heavy Lifter: nnU-Net

To do the actual work of finding the injuries, they used a powerful AI framework called nnU-Net, specifically a "Large" version with a "Residual Encoder." You can think of this as a very strong, flexible net designed to catch all kinds of fish, from tiny minnows (small lesions) to huge sharks (large hematomas). This net is trained to recognize patterns, but it needed the "brain-only" filter to work properly.

The Results: A Balanced Scorecard

When they tested their method on the hidden "final exam" (the test set), the results were impressive, though not perfect.

The Overall Score: They achieved an Overall Dice Coefficient of 0.6305. In the world of this challenge, this placed them at a mean position of 5.5 out of all the teams, making it a highly competitive performance.
The Good News: The AI was incredibly good at identifying healthy brain tissue, scoring a 0.9324 for non-lesion areas. This means it rarely made mistakes about what wasn't damaged.
The Hard Part: Identifying the actual injuries (lesions) was much trickier, with a score of 0.4805. The authors explain that this lower number reflects the sheer difficulty of the task; the injuries are so varied in size and shape that even the best AI struggles to catch every single one.

The paper explicitly notes that while some other teams got slightly better scores at finding the lesions, they often did so by making more mistakes on the healthy tissue (creating "false alarms"). The authors' approach found a sweet spot: it was very careful not to call healthy brain "injured," while still doing a solid job finding the real damage.

What This Means (and What It Doesn't)

The authors suggest that their "brain-only" filter is a powerful tool for handling the messy reality of brain injury scans. However, they are careful not to claim this is a magic cure-all. They admit that finding the tiniest, most diffuse injuries is still a challenge. They don't say this solves the problem of TBI; rather, they show that this specific technique makes the computer's "eyes" sharper when looking at these complex scans.

In short, by teaching the computer to ignore the skull and focus only on the brain's own "lighting," the team built a system that is robust, reliable, and ready to help researchers understand these injuries a little better. It's a step forward, not a finish line, but it's a very strong step in the right direction.

Technical Summary: Lesion Segmentation in Moderate to Severe Traumatic Brain Injury

Problem Statement
The segmentation of lesions in Moderate to Severe Traumatic Brain Injury (msTBI) from T1-weighted MRI presents a significant clinical and computational challenge. Unlike other pathologies such as stroke or tumors, msTBI lesions exhibit profound heterogeneity in size, shape, location, and tissue involvement (spanning gray matter, white matter, and CSF). This variability, combined with the frequent extension of injuries across bilateral regions and multiple tissue types, complicates standard image registration, normalization, and parcellation. Existing automated tools often fail to generalize across these diverse patterns, leading researchers to rely on labor-intensive manual segmentation or to exclude patients with significant injuries, thereby limiting the statistical power of large-scale studies. The AIMS-TBI 2025 Challenge was established to address these barriers by benchmarking robust segmentation algorithms on multi-site, multi-cohort data.

Methodology
The authors propose a deep learning solution based on the nnU-Net framework, specifically utilizing the Residual Encoder Large (resencL) variant. The core innovation lies in the preprocessing pipeline, which introduces an Adaptive Normalization strategy constrained to the brain parenchyma.

Data Source: The study utilized the AIMS-TBI 2025 dataset, comprising 875 T1-weighted MRI scans from 13 international sites (ENIGMA Consortium), covering ages 5–85. The dataset includes diverse injury pathologies (contusions, hematomas, encephalomalacia, etc.) and was split into training (500), validation (100), and test (275) sets.
Anatomically Constrained Adaptive Normalization: Conventional normalization computes mean and standard deviation over the entire image volume, including the skull and background. The authors argue this introduces inter-subject variability due to non-brain structures. Instead, their method:
1. Extracts a brain parenchyma mask using HD-BET.
2. Computes normalization statistics ( $\mu_{brain}$ and $\sigma_{brain}$ ) exclusively within the masked brain region.
3. Normalizes voxel intensities using these brain-specific statistics.
4. This approach aims to reduce inter-subject variance and mitigate artifacts from non-brain structures, allowing the model to focus on pathological regions.
Training Strategy: The model was trained end-to-end using a hybrid loss function (Dice loss + Cross-Entropy). The training incorporated extensive data augmentation (random rotations, scaling, flipping, intensity perturbations, elastic deformations, gamma corrections, and Gaussian noise) to enhance generalization.
Inference: Test-time augmentation (TTA) was applied by mirroring images along spatial axes, with predictions averaged to improve robustness.

Key Contributions

Adaptive Normalization: The primary contribution is the implementation of an intensity normalization strategy restricted to the brain parenchyma. This addresses the specific heterogeneity of msTBI data where non-brain structures can skew global intensity distributions.
nnU-Net Configuration: The application of the resencL (Residual Encoder Large) variant of nnU-Net, which leverages residual blocks to improve feature representation and gradient flow, tailored for the complex lesion patterns of msTBI.
Balanced Performance: The methodology demonstrates a specific trade-off strategy: prioritizing the accurate preservation of non-lesion tissue while maintaining competitive lesion detection, resulting in a robust overall performance metric.

Results
The method was evaluated on the held-out test set of the AIMS-TBI 2025 Challenge using four official metrics: mean position, lesion Dice, non-lesion Dice, and Overall Dice.

Final Performance: The proposed method ("DA5 + Adaptive Norm.") achieved an Overall Dice Coefficient of 0.6305, placing the team at a mean position of 5.5 on the leaderboard.
Component Breakdown:
- Lesion Dice: 0.4805. The authors note this score reflects the inherent difficulty of detecting highly heterogeneous lesions.
- Non-Lesion Dice: 0.9324. This high score indicates the model's strong specificity in correctly identifying non-lesion voxels.
Ablation Studies: Preliminary development results (Table 1) showed that combining extensive data augmentation (DA5) with Adaptive Normalization yielded the highest scores in both development and final testing phases, outperforming strategies using augmentation alone.
Comparison: While some competing teams achieved higher lesion Dice scores, they often did so at the cost of reduced non-lesion accuracy (higher false positives). The authors' approach maintained a well-balanced performance, avoiding over-segmentation.

Significance and Claims
The paper claims that incorporating anatomically constrained normalization within the nnU-Net pipeline is a powerful and effective strategy for tackling the complexities of msTBI lesion segmentation. The results suggest that reducing inter-subject intensity variance by focusing normalization on the brain parenchyma significantly improves model robustness on heterogeneous, multi-site MRI data.

The authors maintain a modest stance regarding the lesion detection capability (Dice ~0.48), acknowledging that detecting small or diffuse lesions remains difficult. However, they emphasize that their approach successfully balances lesion segmentation with the preservation of healthy tissue fidelity. They conclude that accurate automated segmentation is a prerequisite for advanced neuroimaging analyses (e.g., connectivity, outcome prediction) and that future work should focus on enhancing sensitivity to small/diffuse lesions through multi-scale features and multi-modal integration, without compromising the specificity of non-lesion regions.