AIMS-TBI 2025 Challenge

• AIMS-TBI 2025 Challenge is a benchmark for fully automated segmentation of moderate-severe TBI lesions on T1-weighted MRI, addressing diverse lesion morphologies across multiple tissue types.
• The challenge leverages advanced architectures like nnU-Net ensembles and transfer learning from large-scale pretraining to improve segmentation performance, achieving overall Dice scores near 0.60 on unseen test cases.
• Key challenges include detecting small lesions (<100 mm³), reducing false positives with tailored post-processing, and ensuring robust generalization across multisite and demographically diverse datasets.

The AIMS-TBI 2025 Challenge is a benchmark problem in automated neuroimaging analysis centered on fully automated detection and segmentation of moderate-severe traumatic brain injury lesions on T1-weighted 3D MRI. Its central difficulty is extreme lesion heterogeneity: lesions may be focal or diffuse, unilateral or bilateral, and can span volumes from less than $10 \text{ mm}^3$ to more than $10^6 \text{ mm}^3$, while appearing in gray matter, white matter, and the ventricles. This heterogeneity directly affects downstream workflows such as image registration and brain parcellation, so the challenge is not merely a segmentation exercise but a test of whether algorithmic lesion maps are reliable enough for broader neuroimaging pipelines (Dhamat et al., 29 Sep 2025). Closely related work from the AIMS-TBI Segmentation Challenge 2024 established a precursor line of research in which large-scale supervised pretraining on diverse medical segmentation datasets improved msTBI lesion segmentation on T1-weighted MRI by up to 2 Dice points, providing an important technical backdrop for the 2025 competition (Ulrich et al., 9 Apr 2025).

1. Problem definition and challenge scope

The challenge seeks robust, fully automated methods for segmenting moderate-severe TBI lesions from T1-weighted MRI, with explicit attention to lesion size, number, and laterality. The emphasis on T1-weighted imaging is significant because it aligns the benchmark with a modality that is widely used in clinical practice, while also forcing methods to operate under contrast conditions that can be unfavorable for subtle lesion depiction (Dhamat et al., 29 Sep 2025).

A defining property of the benchmark is that lesion morphology is not confined to a narrow phenotype. Participating methods must cope with both large and minute abnormalities, including lesions close to the ventricles and lesions distributed across multiple tissue types. This makes the challenge structurally different from many organ or tumor segmentation tasks in which foreground classes are more spatially constrained. A plausible implication is that architectural choices that improve multi-scale context aggregation, false-positive control, and calibration on no-lesion cases become unusually important.

The 2025 challenge also sits within a broader AIMS-TBI research trajectory. The precursor 2024 segmentation study framed msTBI lesion segmentation as a task in which conventional image processing methods often fail because lesion heterogeneity degrades registration and parcellation. That study proposed a large-scale pretraining strategy inspired by MultiTalent, using a ResencL network pretrained on 48 publicly available segmentation datasets before fine-tuning on msTBI data, thereby establishing transfer learning as a central design pattern for this problem class (Ulrich et al., 9 Apr 2025).

2. Dataset composition and annotation regime

One detailed account of the AIMS-TBI 2025 Challenge describes a dataset of 875 T1-weighted 3D MRI volumes collected across 13 sites, acquired at 1.5 T and 3 T, spanning ages 5–85 years, with 64% male participants. The partition reported for that challenge dataset was 552 training cases, 100 validation cases, and 223 unseen test cases; within the 552 training cases, 528 were used for learning and 24 were held out locally (Dhamat et al., 29 Sep 2025).

The lesion annotations were generated with a three-step, multi-rater, consensus-based manual protocol and were treated as the gold standard. All scans were defaced for privacy. The combination of multisite acquisition, broad age range, mixed scanner field strengths, and consensus annotation gives the challenge a strong emphasis on generalization rather than narrow in-distribution fitting (Dhamat et al., 29 Sep 2025).

Related precursor work used a different msTBI fine-tuning split of 388 training, 100 validation, and 275 held-out test T1-weighted MRI scans, with lesions manually delineated by dual raters and reviewed by experts, and with inter-rater Dice $\ge 0.6$ required for primary raters. That study also resampled all volumes to $1 \text{ mm}^3$ isotropic resolution, applied per-volume intensity Z-score normalization, and used defacing on TBI scans for privacy (Ulrich et al., 9 Apr 2025). The coexistence of these two reported experimental configurations indicates that the AIMS-TBI effort has been accompanied by evolving benchmark formulations and supporting studies rather than a single immutable dataset definition.

3. Methodological landscape

A prominent AIMS-TBI 2025 submission built all segmentation models on the nnU-Net framework, exploiting its automated preprocessing, network configuration, and optimization. Within that framework, three principal components were explored: a 3D UNet baseline, a UNet++ variant with nested dense skip pathways, and a DenseNet-based slice classifier used for post-processing (Dhamat et al., 29 Sep 2025).

The reported 3D UNet used five resolution levels, consisting of four down-sampling stages plus a bottleneck, with feature maps $[32, 64, 128, 256, 512]$. Each level comprised two successive $3 \times 3 \times 3$ convolutions with ReLU and batch normalization, followed by $2 \times 2 \times 2$ max-pooling in the encoder or up-sampling in the decoder, with direct skip concatenations. UNet++ retained the same depth and feature-map counts but inserted additional lateral convolutions to refine semantic fusion across nested skip pathways. The final ensemble combined five probability maps from 5-fold cross-validated UNet models with a single UNet++ probability map according to

$$\bar P_{\text{UNet}}(x)=\frac{1}{5}\sum_{i=1}^{5}P_i(x),$$

and

$$P_{\text{ens}}(x)=\alpha\cdot \bar P_{\text{UNet}}(x)+(1-\alpha)\cdot P_{\text{UNet++}}(x),$$

with $\alpha=0.5$, followed by thresholding at 0.5 to obtain a binary mask (Dhamat et al., 29 Sep 2025).

Preprocessing in that submission included ANTsPyX N4 bias-field correction, automatic resampling to isotropic $10^6 \text{ mm}^3$0 voxels, per-volume z-score normalization, and tight cropping to the brain region using nnU-Net’s connected-component foreground detection. On-the-fly augmentation comprised random flips along all three axes, gamma augmentation with $10^6 \text{ mm}^3$1, random rotations $10^6 \text{ mm}^3$2, and scale and intensity shifts (Dhamat et al., 29 Sep 2025).

The precursor pretraining study pursued a different strategy. Its backbone was a 3D ResencL U-Net from “extending nnU-Net, revisited,” with four encoder levels, a bottleneck at 512 features, four decoder levels, and dataset-specific segmentation heads during pretraining. The pretraining corpus contained 48 publicly available segmentation datasets spanning CT, PET, and multiple MRI sequences, totaling approximately 17,000 scans and covering cardiac structures, abdominal organs and tumors, lung lesions, head-neck OARs and tumors, neuroanatomy, and neuropathology. Pretraining used patch size $10^6 \text{ mm}^3$3, batch size 24, and class-balanced sampling by “inverse $10^6 \text{ mm}^3$4” weighting across datasets before fine-tuning on msTBI-specific data (Ulrich et al., 9 Apr 2025).

Taken together, these approaches define two major methodological directions within the AIMS-TBI ecosystem. One is challenge-oriented architecture ensembling and post-processing within nnU-Net. The other is transfer learning through large-scale supervised pretraining on heterogeneous medical segmentation corpora. This suggests that the competition has functioned not only as a leaderboard benchmark but also as a testing ground for whether anatomical and pathological priors acquired outside TBI transfer effectively into msTBI lesion segmentation.

4. Optimization, losses, and reported performance

The nnU-Net-based 2025 submission optimized a combined Dice and cross-entropy objective per batch. Dice loss was defined as

$10^6 \text{ mm}^3$5

cross-entropy loss as

$10^6 \text{ mm}^3$6

and the total loss as $10^6 \text{ mm}^3$7. Training used SGD with momentum $10^6 \text{ mm}^3$8, weight decay $10^6 \text{ mm}^3$9, initial learning rate $\ge 0.6$0 with linear decay to zero over 1000 epochs, batch size 2 using $\ge 0.6$1 voxel patches, 5-fold cross-validation for UNet, and single-fold training for UNet++ (Dhamat et al., 29 Sep 2025).

On the unseen test set, two settings were emphasized: a UNet plus slice-classifier post-processing pipeline and a UNet plus UNet++ ensemble. Their reported performance was as follows.

Setting	Components	Unseen test performance
Setting 3	UNet + slice-classifier	Accuracy 0.800; DSC-Lesion $\ge 0.6$2; DSC-no $\ge 0.6$3; Overall DSC $\ge 0.6$4
Setting 7	UNet + UNet++ ensemble	Accuracy 0.8451; DSC-Lesion $\ge 0.6$5; DSC-no $\ge 0.6$6; Overall DSC $\ge 0.6$7

Paired $\ge 0.6$8-tests showed that Setting 7 significantly outperformed Setting 3 with $\ge 0.6$9 on accuracy and overall DSC. In the official AIMS-TBI leaderboard, the ensemble ranked among the top-6 methods. The same report states that the overall DSC of approximately 0.60 exceeded the challenge mean of approximately 0.53 and was within 0.03 of the best entry, while exact competitor scores were proprietary (Dhamat et al., 29 Sep 2025).

The precursor pretraining study reported a complementary set of findings. In 5-fold cross-validation on 388 training cases, a baseline 3D U-Net without pretraining achieved $1 \text{ mm}^3$0 Dice, a ResencL U-Net without pretraining achieved $1 \text{ mm}^3$1, ResencL with pretraining and fine-tuning at $1 \text{ mm}^3$2 achieved $1 \text{ mm}^3$3, and ResencL with pretraining plus warm-up to $1 \text{ mm}^3$4 achieved $1 \text{ mm}^3$5, yielding a best-versus-baseline gain of $1 \text{ mm}^3$6 Dice points. On its final test set, using an ensemble of two warm-up models with challenge aggregation, the reported scores were Dice $1 \text{ mm}^3$7 and NSD $1 \text{ mm}^3$8 across all patients, with male $1 \text{ mm}^3$9 and female $[32, 64, 128, 256, 512]$0 (Ulrich et al., 9 Apr 2025).

These results support two non-exclusive observations. First, within a fixed nnU-Net ecosystem, architectural diversity and ensembling can outperform more complicated post-processing pipelines. Second, supervised pretraining on a very broad medical segmentation corpus can produce measurable gains over training from scratch, even when the target pathology is unusually heterogeneous.

5. Failure modes, subgroup effects, and recurrent technical difficulties

The challenge’s primary technical obstacle is the joint effect of extreme heterogeneity and class imbalance. The 2025 submission explicitly noted that many subjects had no lesions or only very small lesions, and that detection of lesions smaller than $[32, 64, 128, 256, 512]$1 remained difficult. In that study, DSC strongly correlated with lesion size, with Pearson $[32, 64, 128, 256, 512]$2, indicating that small-lesion performance is a central limiting factor rather than a marginal edge case (Dhamat et al., 29 Sep 2025).

The same work evaluated two kinds of post-processing beyond plain segmentation. A 2D slice-based false-positive filter used a DenseNet classifier to reject masks when more than 50% of predicted-positive axial slices were classified as “no-lesion” for subjects with predicted lesion volume below $[32, 64, 128, 256, 512]$3. A radiomics-based voxel-wise filter computed 25 top radiomic features with PyRadiomics and classified voxels with XGBoost when predicted lesions were below 1000 voxels. The study’s conclusion was that the simpler UNet plus UNet++ ensemble better balanced sensitivity and specificity than these more elaborate post-processing schemes; the slice classifier removed false positives on no-lesion scans but sometimes removed subtle true lesions under $[32, 64, 128, 256, 512]$4, while the radiomics-based filters did not reliably distinguish small lesions from artifacts and were computationally expensive (Dhamat et al., 29 Sep 2025).

Qualitative analysis in the precursor pretraining study identified a related set of failure modes. Typical moderate msTBI lesions could be segmented with excellent agreement to ground truth, sharp boundaries, and minimal false positives. By contrast, small diffuse lesions near the ventricles were sometimes under-segmented, with occasional false positives in periventricular white matter. Reported failure modes also included missing sub-millimeter micro-hemorrhages and confusion in regions of post-traumatic encephalomalacia or CSF-GM contrast shifts (Ulrich et al., 9 Apr 2025).

Subgroup performance differences were also reported. In the precursor study, male performance exceeded female performance on the final test set, which the authors described as reflecting training set imbalance. This makes demographic imbalance an explicit empirical concern rather than a generic fairness abstraction (Ulrich et al., 9 Apr 2025). A plausible implication is that validation protocols for AIMS-TBI systems need demographic and lesion-size stratification, because aggregate Dice alone can conceal systematic weakness in clinically important subgroups.

6. Related TBI machine learning directions and prospective developments

Two broader TBI machine learning lines are relevant to the future trajectory of the AIMS-TBI challenge. The first is transfer learning with foundation models for ICU time series. One study evaluated the 385 million-parameter encoder-decoder Transformer “MOMENT-1-large,” pretrained by self-supervised masked reconstruction on the Time Series PILE of approximately 13 million univariate series, then fine-tuned for intracranial pressure forecasting using roughly 5,000 hours of TRACK-TBI data. With encoder weights frozen and only a forecasting head trained, the model achieved internal 5-fold cross-validation MAE $[32, 64, 128, 256, 512]$5 mm Hg, compared with $[32, 64, 128, 256, 512]$6 for an LSTM and $[32, 64, 128, 256, 512]$7 for exponential smoothing; on external CHARIS validation, MOMENT achieved MAE $[32, 64, 128, 256, 512]$8, LSTM $[32, 64, 128, 256, 512]$9, and exponential smoothing $3 \times 3 \times 3$0 (Leeuwen et al., 2024). Although this is a forecasting rather than segmentation task, it demonstrates that pretrained representations can be competitive in TBI settings with limited labeled data.

The second line is disease-course modeling from heterogeneous ICU records. A recurrent modeling study on 1,550 moderate-to-severe TBI patients from 65 European ICUs integrated 1,166 pre-ICU and ICU variables into token-embedded two-hour windows and predicted six-month GOSE with an ordinal recurrent model. The full variable set explained up to $3 \times 3 \times 3$1 of ordinal variance, while static pre-ICU and admission information accounted for up to $3 \times 3 \times 3$2 of that explanation; ICU course variables added up to $3 \times 3 \times 3$3 by one week (Bhattacharyay et al., 2023). For AIMS-TBI, this indicates that multimodal TBI modeling may benefit from combining imaging with rich static covariates and carefully selected dynamic signals, rather than assuming that temporal ICU data will automatically dominate the predictive signal.

Specific future directions have been articulated directly in AIMS-TBI-related work. Recommendations from the precursor segmentation study included incorporating additional TBI-specific annotations such as FLAIR and SWI to capture micro-bleeds, balancing training cohorts by gender and age, using synthetic lesion augmentation to simulate hemorrhage and edema, integrating attention gates or transformer blocks in skip connections, exploring multi-scale input branches such as $3 \times 3 \times 3$4 and $3 \times 3 \times 3$5 patches, adding boundary-aware terms or focal-Tversky loss, adopting curriculum learning, using full test-time augmentation with probability ensembling in float32, ensembling diverse schedules for uncertainty estimation, releasing pretrained weights and code recipes, and standardizing aggregation for no-lesion cases (Ulrich et al., 9 Apr 2025).

The 2025 submission proposed a partially overlapping agenda: transformer-based encoders such as Swin-UNet to capture global context and small-object cues, self-supervised pretraining on large unlabeled brain MRI corpora, adaptive loss weighting or focal Dice variants to prioritize small lesion voxels, and multi-task learning with synthetic lesion augmentation (Dhamat et al., 29 Sep 2025). Viewed together, these proposals suggest a convergence toward three priorities for subsequent AIMS-TBI iterations: stronger priors through pretraining, more explicit handling of small and diffuse lesions, and better standardization of evaluation for no-lesion and low-burden cases.

In this sense, the AIMS-TBI 2025 Challenge is best understood as a technically demanding segmentation benchmark at the intersection of heterogeneous lesion morphology, multisite MRI generalization, and transfer-learning-driven model design. Its published results indicate that both architecture ensembling within nnU-Net and large-scale supervised pretraining are viable responses to the core difficulty of msTBI lesion segmentation, while persistent failure on tiny, diffuse, or periventricular lesions continues to define the frontier of the task (Dhamat et al., 29 Sep 2025).