BraTS-Path Challenge 2024
- BraTS-Path Challenge 2024 is a histopathology benchmark that automates the identification of heterogeneous glioblastoma sub-regions from digitized H&E slides.
- It employs a multi-institutional, manually annotated dataset with patch-level classification to address challenges like class imbalance and domain variability.
- The challenge fosters radiology-pathology integration by linking detailed histologic features with clinical imaging and genomic outcomes.
BraTS-Path Challenge 2024 is the first histopathology-focused member of the BraTS Cluster of Challenges and defines a benchmark for automated identification of heterogeneous histopathologic brain tumor sub-regions in glioblastoma from digitized hematoxylin and eosin (H&E) whole-slide images (Bakas et al., 2024). It was introduced to provide a systematically prepared comprehensive dataset and a benchmarking environment for deep-learning models that identify tumor sub-regions of distinct histologic profile, extending BraTS beyond multi-parametric MRI into computational neuropathology and establishing a basis for future radiology-pathology integration (Bakas et al., 2024).
1. Clinical motivation and position within the BraTS ecosystem
Glioblastoma is the most common primary adult brain tumor and has a grim prognosis, with median survival of 12–18 months following treatment and 4 months otherwise (Bakas et al., 2024). A central obstacle in its diagnosis and management is marked histopathologic and micro-environmental heterogeneity. On routine H&E sections, neuropathologists must recognize distinct morpho-pathological features distributed across tissue sections, including cellular tumor, geographic necrosis, pseudopalisading necrosis, areas abundant in microvascular proliferation, infiltration into the cortex, wide extension in subcortical white matter, leptomeningeal infiltration, regions dense with macrophages, and the presence of perivascular or scattered lymphocytes (Bakas et al., 2024).
BraTS-Path 2024 was designed around the proposition that standardized automated detection of these sub-regions can reduce subjectivity, enable large-scale studies connecting histology, imaging, genomics, and outcomes, and assist diagnosis and grading in a consistent manner (Bakas et al., 2024). In the 2024 BraTS cluster, this pathology task complements MRI-centered challenges that use the same broader organizational logic of clinically anchored labels, hidden-set evaluation, and standardized benchmarking (Verdier et al., 2024, LaBella et al., 2024). The challenge description also places BraTS-Path in the same Synapse, MLCommons MedPerf, and GaNDLF ecosystem used across BraTS tracks, which is important for methodological comparability and downstream dissemination (Bakas et al., 2024).
A common misunderstanding is to treat BraTS-Path as merely a pathology analogue of MRI subregion segmentation. The 2024 edition is more specific: it reframes heterogeneous glioblastoma histology as a supervised patch-level classification problem derived from whole-slide region annotations, rather than a native whole-slide segmentation benchmark (Bakas et al., 2024).
2. Dataset construction, annotation protocol, and label taxonomy
The dataset is built from publicly available H&E-stained, formalin-fixed paraffin-embedded digitized tissue sections from the TCGA-GBM and TCGA-LGG collections in The Cancer Imaging Archive, linked to the NCI Genomic Data Commons (Bakas et al., 2024). All cases were re-annotated according to the 2021 WHO CNS tumor classification to identify GBM, IDH-wildtype, WHO grade 4. This included some cases originally in TCGA-LGG that were reclassified as molecular GBM and excluded some cases originally labeled TCGA-GBM when molecular profiles did not meet current WHO GBM criteria (Bakas et al., 2024). For each included case, a single best-quality H&E-stained FFPE tissue section was selected, and frozen-section slides were not used in order to avoid hydration artifacts (Bakas et al., 2024).
The cohort is multi-institutional, drawing from 11 institutions: Henry Ford Hospital, University of California, MD Anderson Cancer Center, Emory University, Mayo Clinic, Thomas Jefferson University, Duke University School of Medicine, Saint Joseph Hospital and Medical Center, Case Western Reserve University, University of North Carolina, and Fondazione IRCCS Istituto Neurologico C. Besta (Bakas et al., 2024). This introduces real-world variability in staining protocols, scanners, and acquisition settings; exact staining and scanner details are not standardized by TCIA or GDC (Bakas et al., 2024).
Ground truth was generated by fully manual whole-slide annotation. A clinically approved protocol defined by experienced neuropathologists specified what each histologic feature should include. Annotators could use any preferred software, and a dedicated Digital Slide Archive instance at Indiana University School of Medicine was provided as a web-based alternative (Bakas et al., 2024). Each case was assigned to one annotator and one approver; the approver was an experienced board-certified neuropathologist with more than 10 years of experience. If annotations were inadequate they were removed, and if accepted annotations produced fewer than approximately 1,500 patches for that slide, the case was returned for additional annotation. The process iterated until annotation quality and patch count were acceptable (Bakas et al., 2024).
After annotation, each region was divided into 512×512 pixel patches. This size was chosen to reduce boundary noise from partial inclusion of regions and to ensure that annotated structures were reasonably contained within a patch. For very small regions of interest, including some microvascular proliferation and lymphocyte-rich regions, curation ensured that at least a 128×128 ROI lay inside each 512×512 patch labeled with that class (Bakas et al., 2024). Patches were then assigned a single dominant class or background, even though the underlying pathology is multi-label at the whole-slide level (Bakas et al., 2024).
The full annotation ontology contains nine histologic patterns plus background, but the 2024 challenge focuses on six classes with sufficient sample size because of strong class imbalance (Bakas et al., 2024).
| Class | Histologic pattern | Patch count |
|---|---|---|
| CT | Cellular Tumor | 43,401 |
| NC | Geographic Necrosis | 24,438 |
| IC | Infiltration into Cortex | 14,500 |
| PN | Pseudopalisading Necrosis | 10,101 |
| WM | Penetration into White Matter | 5,791 |
| MP | Microvascular Proliferation | 5,115 |
The less frequent annotated patterns are Regions Dense with Macrophages (DM), Leptomeningeal Infiltration (LI), and Presence of Lymphocytes (PL), with 1,788, 1,534, and 672 patches respectively; these were not part of the main 2024 classification task (Bakas et al., 2024). Background patches were also defined for patches that did not contain any target ROI, but they were not enumerated in the class-distribution table (Bakas et al., 2024).
The released data pipeline did not apply explicit color normalization or stain augmentation. Additional patch-level artifact filtering was considered, and a three-step artifact removal strategy was described as a recommended approach for future datasets, but it was not applied to the released BraTS-Path 2024 data because the annotation process had already deliberately avoided tissue folding or tearing, pen markings, and glass slippage or reflections (Bakas et al., 2024).
3. Task formulation, infrastructure, and evaluation design
BraTS-Path 2024 defines a patch-level supervised classification task on patches extracted from H&E-stained glioblastoma whole-slide images (Bakas et al., 2024). The input is a 512×512 H&E patch, and the output is a predicted histopathologic class for that patch, or background (Bakas et al., 2024). Although source annotations are region-level segmentations on whole-slide images, each extracted patch is treated as an individual “case,” and each patch receives a single dominant label (Bakas et al., 2024).
The benchmark uses a standard BraTS split structure with training, validation, and test sets; labeled training data are available from Synapse, while validation and test labels are hidden (Bakas et al., 2024). Evaluation is executed through MLCommons MedPerf, which runs participant containers on site data, computes metrics per site, and returns results to Synapse for aggregation and ranking (Bakas et al., 2024). Metrics are implemented in the open-source GaNDLF framework (Bakas et al., 2024).
The challenge uses multiple complementary classification metrics: Accuracy, F1 score, Matthews Correlation Coefficient (MCC), AUROC, Sensitivity, and Specificity (Bakas et al., 2024). Their use reflects the organizers’ explicit concern that strong class imbalance could make raw accuracy misleading. F1 and MCC are therefore central because they better reflect performance under imbalanced label frequencies, while Sensitivity and Specificity expose over-classification and under-classification tendencies (Bakas et al., 2024). The exact macro- or micro-averaging scheme is not specified in the challenge description, and the paper does not state a single primary ranking metric or a weighted composite rule (Bakas et al., 2024).
This evaluation design differs materially from most MRI BraTS tasks. Whereas several BraTS 2024 MRI challenges use lesion-wise Dice and 95th percentile Hausdorff distance as cluster-wide segmentation metrics (Verdier et al., 2024), BraTS-Path 2024 is explicitly a patch-level classification benchmark with classification-style operating characteristics (Bakas et al., 2024). A recurrent misconception is therefore that BraTS-Path 2024 is already a whole-slide segmentation challenge; in fact, whole-slide segmentation and classification are framed as future task evolutions rather than the released 2024 task definition (Bakas et al., 2024).
4. Published challenge solution and reported performance
The BraTS-Path challenge paper itself does not provide organizer baselines or reference performance numbers (Bakas et al., 2024). A separate publication describes a complete BraTS-Path 2024 solution that ranked second in the testing phase and thereby provides the clearest published implementation-level account of how participants approached the benchmark (Zhang et al., 24 Jul 2025).
That solution formulates BraTS-Path as single-instance patch classification with a ResNet-18 backbone initialized from ImageNet weights (Zhang et al., 24 Jul 2025). Patches are resized to 256×256, processed independently, and classified through a final linear layer with a 6-dimensional output corresponding to the six histologic classes. No multiple-instance learning, attention pooling, or slide-level aggregation is used (Zhang et al., 24 Jul 2025). The architecture is therefore a comparatively minimal transfer-learning CNN adapted to six-way histologic subregion recognition.
Training uses weighted cross-entropy to mitigate class imbalance, with OpenCV-based preprocessing that converts BGR to RGB, scales 8-bit intensities to , and applies dataset normalization to zero mean and unit standard deviation (Zhang et al., 24 Jul 2025). The optimization setup is Adam with learning rate , betas , batch size 64, and up to 300 epochs with early stopping (Zhang et al., 24 Jul 2025). Development combined a stratified 80/20 split with 5-fold cross-validation, and final predictions were obtained by averaging the outputs of five fold-specific models (Zhang et al., 24 Jul 2025). All experiments were run in PyTorch on a single NVIDIA RTX 3090 with 24 GB of VRAM (Zhang et al., 24 Jul 2025).
The discrepancy between local and official evaluation was substantial. On a local 20% validation subset drawn from labeled training data, the ensemble achieved Accuracy 0.9872, Recall 0.9872, F1-score 0.9872, Specificity 0.9974, and MCC 0.9828 (Zhang et al., 24 Jul 2025). On the hidden Synapse online validation set, performance fell to Accuracy 0.392229, Recall 0.392229, F1-score 0.392229, Specificity 0.898704, and MCC 0.255267 (Zhang et al., 24 Jul 2025). The paper describes this official validation performance as poor, even though the method placed second in the testing phase (Zhang et al., 24 Jul 2025).
These numbers make BraTS-Path 2024 notable for the gap between conventional internal validation and hidden challenge evaluation. They also show that, at least in one published solution, a comparatively simple transfer-learning classifier could remain competitive in rank despite modest absolute official scores (Zhang et al., 24 Jul 2025).
5. Methodological difficulties and recurrent interpretive issues
BraTS-Path 2024 concentrates several difficulty sources typical of digital pathology but unusual in the original MRI-centered BraTS lineage. First, glioblastoma histology is intrinsically heterogeneous, and many transitions between sub-regions are gradual rather than sharply bounded. Distinguishing cellular tumor from infiltrative cortex or white matter, or separating pseudopalisading necrosis from broader necrotic patterns, is context dependent (Bakas et al., 2024). Because the released task is patch based, wider whole-slide context is absent during classification, which the challenge paper explicitly identifies as a limitation and a reason why multi-scale, context-aware methods are likely to be beneficial in future settings (Bakas et al., 2024).
Second, class imbalance is structural rather than incidental. The main task already excludes LI, DM, and PL because of low sample counts, and even within the six-class benchmark the range from 43,401 CT patches to 5,115 MP patches is large (Bakas et al., 2024). The published second-place solution addressed this only with weighted cross-entropy and fold averaging, and its official validation profile—high specificity but low recall and low MCC—was interpreted as consistent with bias toward majority classes and conservative prediction behavior (Zhang et al., 24 Jul 2025).
Third, domain shift is built into the dataset. Data come from 11 institutions with heterogeneous staining and scanning pipelines, while the released dataset does not impose standardized stain normalization (Bakas et al., 2024). The ResNet-18 solution did not report explicit stain normalization or detailed augmentation, and the large drop from local to hidden validation strongly suggests overfitting and distribution shift between local development data and challenge-held data (Zhang et al., 24 Jul 2025). This suggests that robust stain normalization, stronger augmentation, domain adaptation, or self-supervised representation learning are not optional refinements but central methodological concerns for this benchmark (Bakas et al., 2024).
Fourth, the label structure itself can be misread. The source annotations are region-level and multi-label at the whole-slide scale, but the released challenge labels are patch-level and single-label, with one dominant class per patch (Bakas et al., 2024). A published solution consequently used a six-way classification head without an explicit background output (Zhang et al., 24 Jul 2025). This suggests that practical implementations must reconcile the challenge’s background notion with a six-class operational setup, particularly when training and evaluation pipelines simplify the patch vocabulary.
Inter-rater variability is another unresolved issue rather than a closed matter. Three cases were annotated independently by all annotators to analyze acceptable variation, but detailed agreement statistics were not reported in the challenge paper (Bakas et al., 2024). The benchmark therefore provides expert-reviewed labels while still leaving room for future analysis of annotation uncertainty and label noise.
6. Relation to the wider BraTS program and future development
BraTS-Path 2024 is not an isolated pathology competition but part of a broader attempt to align histology, radiology, and clinically relevant endpoint modeling under a single BraTS framework (Bakas et al., 2024). The challenge description explicitly motivates automated quantification of histologic sub-regions not only for diagnosis and grading but also for linking imaging and biology, including relations such as cellularity versus ADC and microvascular proliferation versus rCBV (Bakas et al., 2024). In that sense, BraTS-Path makes microscopic tumor architecture available to the same ecosystem that already supports MRI segmentation, synthesis, radiogenomics, and outcome modeling.
The broader BraTS software ecosystem had not yet absorbed this pathology branch at the time of the first BraTS orchestrator release. The orchestrator explicitly cites the BraTS-Path challenge but restricts its current scope to MRI-centered tasks, excluding histology because of “the fundamental differences inherent in histological imaging modalities” (Kofler et al., 13 Jun 2025). At the same time, that paper presents BraTS-Path as part of the same challenge family and describes how a pathology-oriented extension would require a new task type, WSI loading, color normalization, tiling or patch extraction, histology-specific post-processing, and centralized metric handling (Kofler et al., 13 Jun 2025). This suggests that BraTS-Path is institutionally integrated into BraTS even where implementation infrastructure remains MRI specific.
Future directions stated across the challenge and solution papers are relatively consistent. The challenge paper points toward expansion from patch-level classification to whole-slide segmentation and classification, denser annotations, incorporation of rarer classes into primary tasks, and explicit analysis of inter-rater variability (Bakas et al., 2024). The second-place solution paper points toward stronger generalization across institutions and data sources, more powerful backbones, multi-scale architectures, multiple-instance learning, and more careful case- or slide-level validation strategies (Zhang et al., 24 Jul 2025). A plausible implication is that later BraTS-Path iterations will move from patch-wise recognition of dominant morphology toward genuinely hierarchical neuropathology modeling: whole-slide spatial reasoning, uncertainty-aware supervision, and tighter coupling to MRI-defined tumor phenotypes.
In that broader historical sense, BraTS-Path Challenge 2024 established a standardized pathology benchmark inside a community previously dominated by volumetric MRI segmentation. Its distinctiveness lies not simply in using digitized slides, but in formalizing heterogeneous glioblastoma histology as a reproducible computational task, with expert-reviewed annotations, multi-institutional heterogeneity, containerized evaluation infrastructure, and an explicit route toward radiology-pathology integration (Bakas et al., 2024).
