BraTS-Lighthouse 2025 Challenge

• The BraTS-Lighthouse 2025 Challenge is a multi-task ecosystem that extends beyond single glioma segmentation to cover diverse tumor types and clinically realistic imaging tasks.
• It employs a rigorous 4-stage annotation pipeline combining manual and AI-assisted segmentation to evaluate intra- and inter-annotator variability for reliable reference standards.
• The initiative uniquely merges neuroradiology education with algorithm development through interactive lectures, workshops, and one-on-one guided annotation sessions.

The BraTS-Lighthouse 2025 Challenge denotes a 2025 BraTS/MICCAI challenge ecosystem centered on clinically grounded brain tumor imaging tasks, high-quality reference-standard MRI annotation, and explicit integration of neuroradiology education with algorithm development. In the 2025 literature, it is described both as a landmark initiative to develop accurate brain tumor segmentation algorithms and as a broader benchmark spanning segmentation, synthesis, inpainting, and longitudinal response assessment across multiple tumor types and clinical scenarios, including adult glioma, pediatric tumors, meningioma, brain metastases, and related generative tasks (Amiruddin et al., 21 Sep 2025, Capellán-Martín et al., 16 Dec 2025, Maleki et al., 16 Apr 2025).

1. Scope, nomenclature, and challenge landscape

BraTS refers to the Brain Tumor Segmentation Challenge, ASNR to the American Society of Neuroradiology, and MICCAI to the Medical Image Computing and Computer Assisted Interventions Society (Amiruddin et al., 21 Sep 2025). Within the 2025 papers, BraTS-Lighthouse is consistently associated with a shift away from a single glioma-only segmentation benchmark toward a family of clinically differentiated tasks. One challenge paper presents the 2025 benchmark in terms of PED, MEN, MEN-RT, and MET cohorts, each with distinct label structures and imaging assumptions (Capellán-Martín et al., 16 Dec 2025). Other 2025 BraTS papers describe Adult Glioma Segmentation as Task 1, a Brain Tumor Progression task as Task 11, an inpainting track, and a missing-modality synthesis setting linked to BraTS-Lighthouse 2025 or BraSyn 2025 (Jain et al., 29 Sep 2025, Tikhonov et al., 8 Sep 2025, Zhang et al., 25 Nov 2025, Sandbhor et al., 19 Sep 2025).

This suggests that BraTS-Lighthouse 2025 functions as an umbrella rather than a single narrow benchmark. Its common thread is not one anatomy or one output type, but the use of multi-parametric MRI, expert-approved labels or surrogates, and challenge structures designed to approximate real clinical problems such as post-treatment follow-up, longitudinal response assessment, missing acquisition channels, and dataset bias from pathology.

The challenge family is also explicitly multi-institutional. For MET, the reported data sources include Duke University, Washington University in St. Louis, University of Missouri, Yale University, UCSF, UCSD, Northwestern University, and the National Cancer Institute in Egypt (Maleki et al., 16 Apr 2025). For missing-modality synthesis, the retrospective dataset is drawn from BraTS-GLI 2023, BraTS-METS 2023, and BraTS-MENINGIOMA (Sandbhor et al., 19 Sep 2025). The broader direction is therefore one of heterogeneous, clinically realistic neuro-oncology benchmarking rather than narrowly curated single-site evaluation.

2. Reference-standard generation and annotation methodology

A defining feature of BraTS-Lighthouse 2025 is the centrality of reference-standard construction. The educational platform paper describes a 4-stage annotation pipeline with 7-day intervals between stages: manual annotation from scratch, repeated manual annotation from scratch, review and refinement of automated nnU-Net pre-segmented images, and repeated review and refinement of automated nnU-Net pre-segmented images (Amiruddin et al., 21 Sep 2025). The stated purpose of this design was to suppress image recall, evaluate intra- and inter-annotator variability, and support creation of a more reliable reference standard. Annotation used ITK-SNAP version 4.2.0 and 3D Slicer version 5.6.2, with one-on-one meetings, video tutorials, best-practice documents, and guidance to identify and label tumor subregions (Amiruddin et al., 21 Sep 2025).

The BraTS-METS 2025 challenge paper makes this variability objective even more explicit. It introduces a 75-subject multi-annotated dataset in which each case receives four neuroradiologist segmentation instances: two from-scratch segmentations and two segmentations after AI pre-segmentation, performed while recorded on video (Maleki et al., 16 Apr 2025). The challenge frames this as a mechanism for establishing inter-rater and intra-rater variability rather than treating expert annotation as a single unquestioned ground truth.

The 2025 effort also builds on earlier BraTS data-production pipelines. For the 2023 and 2024 MET datasets, the reported workflow was AI pre-segmentation, student annotation or refinement, board-certified neuroradiologist review, quality control, a second neuroradiologist review, and final senior neuroradiologist approval (Maleki et al., 16 Apr 2025). BraTS-Lighthouse 2025 therefore preserves the staged-review logic of earlier BraTS efforts while adding repeated expert annotation as a first-class methodological object.

The annotation workload reported for the educational Lighthouse platform is substantial. Fourteen faculty-coordinator pairs completed 1200 segmentations in total and averaged 1322.9±760.7 hours per dataset per pair (Amiruddin et al., 21 Sep 2025). The tumor datasets annotated were brain metastases, gliomas, untreated meningiomas, and post-treatment meningiomas, with six pairs assigned to brain metastases, two to gliomas, three to untreated meningiomas, and three to post-treatment meningiomas (Amiruddin et al., 21 Sep 2025).

An important manuscript-level caveat appears in the BraTS-METS 2025 paper: the reported total MET case counts are internally inconsistent. The paper states 1778 multi-parametric MRI cases overall, also notes a release of 1475 cases on Synapse with 1296 training and 179 validation cases, and elsewhere describes a split of 1296 training, 179 validation, and 303 testing cases (Maleki et al., 16 Apr 2025). The discrepancy is explicitly present in the manuscript summary and should be treated as such.

3. Educational platform and physician training

BraTS-Lighthouse 2025 is unusual among challenge frameworks in that education is not described as peripheral. The ASNR MICCAI BraTS 2025 Lighthouse Challenge education platform was designed as a dual-purpose initiative: creation of high-quality reference-standard annotated MRI datasets for algorithm development, and a hands-on neuroradiology plus AI education platform for medical students and radiology trainees (Amiruddin et al., 21 Sep 2025).

The platform was intentionally multimodal. It combined faculty-led didactics, lectures, journal clubs, workshops led by data scientists, one-on-one guided annotation sessions, bite-sized social media learning, and asynchronous YouTube access (Amiruddin et al., 21 Sep 2025). The stated design principles were interactive learning, mentorship, adult learning principles, real-world application, and interdisciplinary collaboration among neuroradiologists, students and trainees, and data scientists (Amiruddin et al., 21 Sep 2025).

Participation was substantial. Fifty-six annotators volunteered in BraTS 2023 and 2024 activities, and for BraTS 2025 Lighthouse there were 14 annotation coordinators, each paired one-on-one with a board-certified neuroradiology faculty annotator (Amiruddin et al., 21 Sep 2025). The lecture program covered clinically relevant MRI neuroanatomy, fundamentals of MRI, standardized reporting for brain tumors, molecular imaging of brain tumors, pediatric brain tumors, skull base imaging, fundamentals of artificial intelligence, AI in brain tumor imaging, and AI implementation in low-resource settings (Amiruddin et al., 21 Sep 2025).

The educational outcomes are reported quantitatively. On a 1–10 scale, familiarity with image segmentation software increased from 6 ± 2.9 to 8.9 ± 1.1, and familiarity with brain tumor features increased from 6.2 ± 2.4 to 8.1 ± 1.2, with Student’s $t$ test $p<0.05$ for pre/post improvements (Amiruddin et al., 21 Sep 2025). Among lecture attendees, 95% (92/97) reported improved knowledge (Amiruddin et al., 21 Sep 2025). In a survey of 54 medical students about AI education needs, 93% (50/54) believed AI would influence their careers, 87% (47/54) reported no AI-focused education at their institutions, and 72% (39/54) considered AI-focused training important (Amiruddin et al., 21 Sep 2025).

A common misconception is that BraTS-Lighthouse 2025 is solely a leaderboard exercise. The educational platform paper directly contradicts that reading: learning was integrated into the actual segmentation workflow so that trainees could see how clinical labeling, quality control, and AI model development connect (Amiruddin et al., 21 Sep 2025).

4. Task families, cohorts, and label systems

The 2025 papers collectively describe a heterogeneous task landscape.

Track or cohort	Objective	Labels or outputs
Adult Glioma Segmentation	Pre-treatment and post-treatment segmentation	NETC, SNFH, ET, RC; TC, WT
PED	Pediatric tumor segmentation	ET, NET, CC, ED; TC, WT
MEN / MEN-RT	Preoperative or radiotherapy segmentation	ET, NETC, SNFH; or GTV
MET	Pre-treatment and post-treatment metastasis segmentation	ET, NETC, SNFH, RC; TC, WT
Inpainting	Healthy tissue synthesis	t1n infilling image
Missing modality synthesis	Reconstruct absent modality	T1w, T1ce, T2w, FLAIR
Brain Tumor Progression	Longitudinal response prediction	CR, PR, SD, PD

For Adult Glioma Segmentation, the reported subregions are NETC, SNFH, ET, and RC, with TC = NETC + ET and WT = NETC + SNFH + ET (Jain et al., 29 Sep 2025). The key clinical distinction is that RC occurs only in post-treatment scans. For PED, the reported labels are ET, NET, CC, and ED, with TC = ET + NET + CC and WT = TC + ED (Capellán-Martín et al., 16 Dec 2025, Mulvany et al., 2024). For MEN, the labels are ET, NETC, and SNFH, while MEN-RT uses only GTV and only T1CE in original image space (Capellán-Martín et al., 16 Dec 2025). For MET, the labels are ET, NETC, SNFH, and RC, again with TC and WT as composite regions (Capellán-Martín et al., 16 Dec 2025, Maleki et al., 16 Apr 2025).

Non-segmentation tracks broaden the challenge substantially. The inpainting task is defined as synthesis of subject-specific healthy anatomical proxies from pathological MRI, motivated by pathology bias and the lack of pre-disease baseline scans (Zhang et al., 25 Nov 2025). The missing-modality synthesis setting requires generation of any one of T1w, T1ce, T2w, or FLAIR from the remaining modalities (Sandbhor et al., 19 Sep 2025). Task 11 addresses treatment-response assessment in glioblastoma from longitudinal MRI, with prediction of Complete Response, Partial Response, Stable Disease, and Progressive Disease (Tikhonov et al., 8 Sep 2025).

This task diversity explains why several 2025 papers emphasize that tumor types differ widely in anatomy, appearance, and label structure, and why adaptable or task-specific pipelines are repeatedly favored over a single universal architecture (Capellán-Martín et al., 16 Dec 2025).

5. Representative methods and competitive submissions

A recurrent pattern across BraTS-Lighthouse 2025 submissions is that performance gains are attributed less to entirely new backbones than to pipeline design, augmentation, ensembling, and task-specific decomposition.

For Adult Glioma Segmentation, one first-place Task 1 solution used on-the-fly GAN-based augmentation integrated into nnU-Net training. It dynamically inserted synthetic tumors using pretrained GliGANs during training and ensembled a baseline model, a regular on-the-fly augmented model, and a customized on-the-fly augmented model (Jain et al., 29 Sep 2025). On the online BraTS 2025 validation platform, the final ensemble reported lesion-wise Dice of 0.790 for ET, 0.749 for NETC, 0.872 for RC, 0.825 for SNFH, 0.790 for TC, and 0.880 for WT, and the paper states that the method ranked first in BraTS Lighthouse Challenge 2025 Task 1 (Jain et al., 29 Sep 2025).

For the inpainting track, the reported winning method used a 3D U-Net architecture with random masking augmentation and a composite MAE + SSIM loss (Zhang et al., 25 Nov 2025). On the final online test set it achieved SSIM 0.91928125, PSNR 26.9321548, and RMSE 0.05162604, securing first place in the BraTS-Inpainting 2025 challenge (Zhang et al., 25 Nov 2025).

For missing-modality synthesis, SLaM-DiMM combined a shared-latent diffusion-based 2D synthesis stage with a 3D coherence enhancement module based on 3D-UNETR (Sandbhor et al., 19 Sep 2025). Reported validation SSIM values ranged from 94.96 for glioma T1w synthesis to 89.41 for metastasis FLAIR synthesis, with the 3D refinement stage mainly aimed at removing inter-slice discontinuities and vertical blurry artifacts (Sandbhor et al., 19 Sep 2025).

For multi-task segmentation across PED, MEN, MEN-RT, and MET, an adaptable pipeline used radiomic-guided subtype detection, multiple strong backbones, lesion-wise model weighting, and adaptive post-processing (Capellán-Martín et al., 16 Dec 2025). Its ensemble weights were defined by

$W_i = \frac{\sum_{j \neq i} F_j}{\sum_i F_i},$

where $F$ is a rank-based internal score aligned with BraTS ranking logic and smaller is better (Capellán-Martín et al., 16 Dec 2025).

Pediatric BraTS-family work further illustrates the continued importance of nnU-Net-derived baselines. A radiologically informed two-stage cascaded nnU-Net for BraTS-PEDs 2024 used modality-specific refinement branches and reported mean Dice scores of 0.657, 0.904, 0.703, and 0.967 for ET, NET, CC, and ED, respectively (Mulvany et al., 2024). This is not a BraTS-Lighthouse 2025 paper, but it is directly relevant to the same challenge family and label structure.

6. Evaluation logic, clinical orientation, and open issues

BraTS-Lighthouse 2025 evaluation is consistently lesion-aware and clinically motivated. Across the papers, reported metrics include Dice Similarity Coefficient, lesion-wise Dice, Normalized Surface Dice or Normalized Surface Distance, HD95, Sensitivity, Specificity, and Precision (Jain et al., 29 Sep 2025, Maleki et al., 16 Apr 2025, Adewole et al., 2023). MET ranking is described as the sum of ranks across all metrics, with statistical significance testing used for the ranking strategy (Maleki et al., 16 Apr 2025). Several submission papers explicitly mirror this logic internally rather than optimizing a single overlap score (Capellán-Martín et al., 16 Dec 2025).

The clinical orientation of the benchmark is equally explicit. Adult glioma Task 1 emphasizes robust segmentation across both pre-treatment and post-treatment scans for treatment planning, post-surgical monitoring, and outcome prediction (Jain et al., 29 Sep 2025). MET emphasizes lesion detection, volumetric quantification, and treatment-response assessment in longitudinal and post-SRS settings (Maleki et al., 16 Apr 2025). The progression task seeks automated RANO-class prediction from consecutive scan pairs, reporting mean ROC AUC 0.81 and Macro F1 0.50 for a hybrid deep learning and radiomics approach on 616 longitudinal MRI scans from 91 glioblastoma patients (Tikhonov et al., 8 Sep 2025). The inpainting and missing-modality tracks extend the benchmark beyond segmentation to problems of pathology bias, counterfactual healthy-tissue synthesis, and incomplete acquisition protocols (Zhang et al., 25 Nov 2025, Sandbhor et al., 19 Sep 2025).

Two broader issues run through the 2025 challenge literature. First, reference standards are treated as variable rather than absolute; repeated expert annotation and explicit intra-rater and inter-rater analysis are part of the challenge design itself (Maleki et al., 16 Apr 2025). Second, generalization across institutions, protocols, and underrepresented populations remains a central concern. Earlier BraTS-Africa work had already framed BraTS as needing to address domain shift, lower-quality MRI, and equitable deployment in Sub-Saharan Africa (Adewole et al., 2023), and this concern persists in 2025 SSA-focused submissions that emphasize segmentation-aware augmentation and model ensembling on small, heterogeneous, population-specific data (Ankomah et al., 3 Oct 2025).

Taken together, the 2025 BraTS-Lighthouse papers portray a challenge ecosystem in which dataset curation, annotation methodology, educational design, lesion-wise evaluation, and clinically realistic task formulation are all treated as core scientific problems. A plausible implication is that BraTS-Lighthouse 2025 marks a transition from benchmark construction around a single canonical segmentation task toward benchmark construction around the full workflow of AI-assisted clinical neuroradiology.