GLI-AL: A Multi-Modal Glioma MRI Label Resource with Unified Anatomy-Lesion Labels

arXiv:2607.22135 · cs.CV · Submitted 2026-07-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "GLI-AL: A Multi-Modal Glioma MRI Label Resource with Unified Anatomy-Lesion Labels".

Jane: BraTS datasets provide multi-center, pre-operative, multi-parametric MRI and expert tumor-subregion annotations for brain tumor segmentation research, but they are among the central public benchmarks for machine learning in glioma imaging.

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So we're diving into this paper today which is called "GLI-AL: A Multi-Modal Glioma MRI Label Resource with Unified Anatomy-Lesion Labels," and it sounds like they are tackling a really tricky problem in medical imaging. Jane, can you give us the quick rundown on what the main point of this paper is?

Jane: Sure thing, Tom. Basically, the authors address a big issue where existing BraTS datasets focus only on tumor subregions and completely overlook coexisting white matter hyperintensities or WMH in brain MRIs. They explain that when you try to train models for joint segmentation involving WMH, treating those unlabeled abnormalities as normal tissue creates task-specific label noise.

Lu: That is a crucial point because it means the original BraTS-GLI tumor subregion labels simply don't work well as a joint supervision target when you need to include both healthy tissue and these common coexisting abnormalities in the same label space.

Meng: From an engineering standpoint, that sounds like a serious headache for model training pipelines if we don't handle it correctly, because misclassifying pathology as normal tissue definitely messes up the learning signal.

Lalam: I see how this paper is aiming to build a solution that unifies these different pathological and healthy structures into one consistent label set so the AI models don't get confused by those noisy inputs.

Tom: Exactly, and what they claim is that they introduce GLI-AL, which provides one thousand two hundred fifty-one unified eight-class anatomy-lesion label sets aligned with the original four-modal MRI cases from the BraTS two thousand twenty-three-GLI training cohort.

Jane: That unification is achieved through a specific integer coding system where they define seven anatomical classes: zero for background, one for cortical gray matter, two for basal ganglia, three for white matter, four for lesion, five for ventricle, six for cerebellum, and seven for brainstem.

Lu: The idea of merging healthy tissue labels and lesion structures into a single supervision target is smart because it removes the need to switch between incompatible label definitions during downstream model training.

Meng: I'm interested in how they manage the construction of this unified space, especially since they are dealing with unlabeled abnormalities like WMH that need careful handling.

Lalam: The construction workflow involves three main steps: first, purifying a subset by identifying cases with expert negative WMH conclusions and those screened by models; second, extending labels for the remaining cases using tools like DeepWMH and LST-AI to find coexisting abnormalities via intersection; and finally, obtaining healthy tissue labels through a process involving TumorSynth probability maps.

Tom: That sounds like a very detailed approach to data creation, especially how they handle the noise by intersecting candidate abnormalities with the existing BraTS lesion mask. Jane, what does this unified space actually allow users to do that wasn't possible before?

Jane: Users can retain the original BraTS tumor-subregion masks while simultaneously identifying newly added lesion component voxels within that unified Lesion class that fall outside the original whole-tumor mask.

Paper summary: Lu: It’s powerful because it lets researchers keep their established tumor annotations while gaining this richer, more complete understanding of all anatomical structures present in the scan.

Meng: If you look at the data construction, I wonder about the quality control aspect; how do they ensure those probability maps generated by TumorSynth are reliable before fusing them into that unified label?

Lalam: They use an automatic outlier detection based on the interquartile range or IQR to remove low-quality modality probability maps before they are fused together, which helps keep the resulting labels robust.

Tom: That sounds like a solid plan for building this resource, and it leads us right into the big picture of what this resource means for future segmentation research. Lu, you mentioned creativity earlier; what wild possibilities does this unification open up in terms of new types of AI applications?

Lu: I think the ability to train models on such a comprehensive anatomy-lesion map suggests we could develop AI systems that can perform much more nuanced functional assessments based on precise structural context, perhaps linking specific WMH patterns directly to cognitive decline markers.

Jane: That moves us beyond just finding tumors and starts looking at how the entire brain structure, including its common coexisting conditions like WMH, influences function.

Meng: From a practical perspective, if we can use these unified labels reliably across different MRI modalities without needing extensive re-registration for every input case, that drastically simplifies the deployment of these segmentation models in real-world clinical settings.

Lalam: This resource is significant because it essentially cleans up the label space for joint segmentation tasks, which means models trained on this data should exhibit better performance when tested on more complex scenarios involving comorbid conditions.

Tom: So, to wrap up what we've heard about "GLI-AL: A Multi-Modal Glioma MRI Label Resource with Unified Anatomy-Lesion Labels," it’s a resource built from the BraTS two thousand twenty-three-GLI training cohort that systematically adds WMH representation to create a unified label space of one thousand two hundred fifty-one sets.

Jane: The title and authors of this paper point toward a focus on creating a controlled access, labels-only derived resource specifically for glioma MRI segmentation research.

Lu: The implications are that AI can start performing more holistic brain analysis rather than just localized tumor detection because the input data representation is much more complete.

Meng: The impact could be in making diagnostic pipelines more robust by ensuring models don't get confused by common, unlabeled pathology like WMH, which would improve the practical adoption of these tools.

Lalam: This paper's contribution is providing a WMH-aware cleaned training reference for the joint label space, which is vital because it shows that model performance on outof-domain healthy anatomical structure segmentation is preserved when using this resource.

Tom: That preservation of performance across different data domains, especially concerning healthy tissue classes, really validates the effort put into making this unified label set. It suggests we have a more reliable foundation for building models that understand the whole picture.

Conclusion: Tom: So we've been diving deep into this resource, GLI-AL, which is essentially taking existing brain tumor data and adding a whole new layer of detail to make it much more useful for advanced AI segmentation.

Jane: That’s right, Tom; the authors have put together this system to fix a problem where models often get confused by things like white matter hyperintensities when they're trying to segment tumors.

Lu: The title itself tells you everything we need to know about the core innovation here, focusing on that multi-modal aspect and the unified label space.

Meng: I think the authors did a really solid job of taking a messy set of data and making it clean enough for serious engineering work without introducing too much noise during the process.

Lalam: From my perspective, this resource fundamentally improves how we can train AI to understand brain structures holistically, which could lead to much more nuanced functional mapping in the future.

Tom: Exactly, and when you look at who wrote this—the authors—you realize they're tackling one of the most persistent headaches in medical imaging research right now.

Jane: They are clearly focused on creating a controlled way to represent both healthy tissue and lesions within the same mathematical framework for segmentation tasks.

Lu: The real impact, I see it as allowing AI to move beyond just identifying a tumor boundary and start understanding the entire anatomical context around it.

Meng: Practically speaking, this means we can build more reliable AI tools that don't fail when they encounter common co-occurring conditions in patient scans.

Lalam: This work has implications for how we develop medical AI culture because it pushes us toward building systems that are robust against the complex realities of human brain pathology.

Xingyu Xiang, Shuang Hao, Fan Wang, Jianhua Ma, Chunfeng Lian

Key Laboratory of Biomedical Information Engineering of Ministry of Education, School of Life Science and Technology, Xi’an Jiaotong University

cs.CV

Submitted: 2026-07-24

Updated: 2026-07-27

Comments: Minor revision: Figure 1 was repositioned. The scientific content remains unchanged

Journal ref: Machine Learning for Biomedical Imaging 2026 (2026) 834-843

DOI: 10.59275/j.melba.2026-575f

Code: https://github.com/xyx200/brats-gli-anatomy-lesion-code

Project page: https://www.synapse.org/Synapse:syn51156910/wiki/621282

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 86/100

The gist: BraTS datasets provide multi-center, pre-operative, multi-parametric MRI and expert tumor-subregion annotations for brain tumor segmentation research, but they are among the central public benchmarks

Key concepts

Unified Label Space
This is a single integer coding system (0-7) that combines healthy brain tissues and lesion structures into one set of labels. Instead of separate labels for healthy tissue and lesions, every voxel is assigned one code. This simplifies model training because the AI doesn't need to learn two different label definitions, ensuring consistency across all anatomical structures.
Label Noise Mitigation
Existing datasets suffer from 'task-specific label noise' where unlabeled abnormalities like WMH are treated as normal tissue during segmentation. GLI-AL mitigates this by explicitly incorporating WMH information into the unified labels. This ensures that models learn to distinguish between actual lesions and coexisting white matter changes, leading to more robust and accurate segmentation.
Data Construction Workflow
The resource is built in three steps: first, purifying subsets based on expert negative WMH conclusions; second, using deep learning models (DeepWMH/LST-AI) to find potential coexisting abnormalities and merging them with the original lesion masks; and third, generating healthy tissue labels using modality probability maps. This systematic process ensures high quality and relevance for the final unified label set.

Terminology

Summary

BraTS datasets provide multi-center, pre-operative, multi-parametric MRI and expert tumor-subregion annotations for brain tumor segmentation research, but they are among the central public benchmarks for machine learning in glioma imaging. The gist is that this resource introduces a controlled access, labels-only derived resource built from the BraTS 2023-GLI training cohort to address label noise caused by coexisting white matter hyperintensities (WMH) by providing unified eight-class anatomy-lesion labels aligned with original four-modal MRI cases.

Resource Gap and Motivation

Existing BraTS datasets focus on tumor subregions but do not systematically represent coexisting white matter hyperintensities (WMH), which are common in glioma MRI. In joint segmentation settings, unlabeled abnormalities like WMH introduce task-specific label noise by treating pathological regions as normal tissue. This creates a resource gap because the original BraTS-GLI tumor-subregion labels cannot directly serve as a joint supervision target when healthy brain tissues and coexisting abnormalities need to be represented in the same label space. The paper addresses this by introducing GLI-AL, which provides 1,251 unified eight-class anatomy-lesion label sets aligned with the original four-modal MRI cases.

Label Space Unification

The resource establishes a unified label space covering both healthy tissues and lesion structures. This unification is achieved through a specific integer coding system: 0 for background, 1 for cortical gray matter (GM), 2 for basal ganglia (BG), 3 for white matter (WM), 4 for Lesion, 5 for ventricle (Ven), 6 for cerebellum (Cer), and 7 for brainstem (BS). This coding places lesion labels and healthy tissue structures into a single supervision target, meaning downstream models no longer need to switch between incompatible label definitions. Furthermore, users can retain the original BraTS tumor-subregion masks while identifying newly added lesion component voxels within the unified Lesion class that lie outside the original whole-tumor mask.

Data Construction Workflow

The construction of GLI-AL involves three main steps:

  1. Purified subset construction: This involved establishing WMH-aware stratification across all cases and distinguish purified and extended subsets. The purified subset includes cases with expert negative WMH conclusions from the Rudie et al. expert WMH annotation resource and additional WMH-negative cases identified by model screening.

  2. Extended subset label completion: For the remaining 857-case extended subset, candidates for coexisting abnormalities are generated using DeepWMH (Liu et al., 2024) and LST-AI (Wiltgen et al., 2024) to reduce false positives through intersection. This intersection mask is then united with the existing BraTS lesion mask to form the unified Lesion class.

  3. Unified anatomy-lesion labels: Healthy-tissue labels are obtained by running TumorSynth (Wu et al., 2026) on the four MRI modalities to obtain voxel-level softmax probability maps for each modality. An automatic outlier detection based on the interquartile range (IQR) is used to remove low-quality modality probability maps before fusion.

Resource Integrity and Metadata

To ensure reproducibility and scientific rigor, the resource provides comprehensive metadata. The release includes case-level metadata covering label source, image-repair requirements, quality-control status, access conditions, checksums, and release boundaries. This organization allows users to select cases by source and report training or evaluation data use without treating all derived labels as homogeneous samples. Additionally, the resource includes 116 cases providing repair labels with two foreground classes, representing whole tumor and WMH.

Validation and Performance

The resource's effectiveness is validated through a controlled ML use case comparing Baseline and Joint-Baseline models on in-domain GLI data. The results show that the joint model maintains stable healthy-brain-tissue segmentation ability, while its performance on the main lesion is comparable to the baseline. Crucially, in external zero-shot WMH evaluation, the two models still maintain similar results for healthy tissue classes, indicating that model performance on outof-domain healthy anatomical-structure segmentation is preserved. This demonstrates that the resource successfully provides a WMH-aware cleaned training reference for the joint label space.

Responsible Use and Availability

GLI-AL is intended only for scientific research and should not be used for direct clinical diagnosis or treatment decisions. Users must report the data tier and label source used, and state limitations of automatically generated labels. Access requires obtaining upstream BraTS 2023 data access first, followed by applying for access to GLI-AL through a specific email workflow. The resource is released under Creative Commons AttributionNonCommercial 4.0 International (CC BY-NC 4.0).

Improvements for AI systems

Here are the specific improvements that can be made to existing AI systems by utilizing the GLI-AL resource, along with what these improved systems will be capable of doing:

The GLI-AL resource enables significant advancements in medical image segmentation and joint modeling by providing a unified, auditable label space encompassing healthy tissues and comorbid lesions. Here are the specific improvements for AI systems:

  1. Enhancement of Joint Segmentation Models (e.g., MedNeXt, nnUNet):

  2. Improved Robustness to Label Noise from Coexisting Abnormalities:

  3. Enabling Joint Supervision of Healthy Tissue and Lesion Structures:

  4. Facilitating Data-Source Stratified Training and Evaluation:

Specific capabilities of the improved AI systems:

  1. Improved Robustness to Label Noise from Coexisting Abnormalities:

  2. The system can be trained using the unified eight-class label space (including healthy tissues like GM, WM, BG) rather than relying solely on original tumor subregion masks. This directly addresses label noise where unlabeled coexisting white matter hyperintensities (WMH) are implicitly treated as normal tissue during training. The resulting models will exhibit higher sensitivity to true pathological regions by explicitly learning the distinction between healthy anatomy and coexisting lesions, leading to more accurate segmentation in complex, real-world clinical scenarios.

  3. Enabling Joint Supervision of Healthy Tissue and Lesion Structures:

  4. The system can perform simultaneous joint segmentation of normal brain anatomy (e.g., cortical gray matter, white matter) alongside pathological structures (lesions). This eliminates the need for models to switch between incompatible task labels, resulting in a single, coherent supervision target that models can learn from directly. This allows for more holistic understanding of the patient's brain pathology rather than treating lesions and healthy tissue as separate segmentation problems.

  5. Facilitating Data-Source Stratified Training and Evaluation:

  6. Researchers can conduct stratified training by selecting cases based on label source (e.g., purified subset vs. extended subset) or case stratification (purified vs. extended). This allows researchers to rigorously quantify the impact of data quality, sample size, and specific lesion constraints on model performance (as validated in Section 6), enabling the development of more reliable model-ready datasets and more precise analysis of data artifacts.

  7. Improved Reproducibility and Auditability:

  8. The system's training and evaluation can be linked to specific label sources, quality control statuses, and case-level provenance records (including image repair requirements). This provides transparent metadata for every prediction, allowing researchers to trace performance back to the exact data source and processing pipeline used in training or testing, which is crucial for scientific reproducibility.

  9. Enabling Image Repair Reproduction:

  10. The system can leverage the provided image-repair labels (for 116 cases) to reproduce specific repaired imaging inputs. This capability allows researchers to test how model performance changes when input data is corrected for known artifacts, providing a pathway for iterative refinement of both the AI model and the preprocessing pipeline itself.

Abstract

Existing BraTS-GLI datasets provide a widely used benchmark for adult glioma MRI segmentation, but their task definition focuses on tumor subregions and does not systematically represent coexisting white matter hyperintensities (WMH). In joint segmentation settings, such unlabeled abnormalities introduce task-specific label noise by treating pathological regions as normal tissue. To address this limitation, we introduce BraTS-GLI Anatomy-Lesion, a controlled-access, labels-only derived resource built from the BraTS 2023-GLI training cohort. The resource provides 1,251 unified eight-class anatomy-lesion label sets aligned with the original four-modal MRI cases, including image-repair labels for 116 cases requiring repaired imaging inputs. The cohort is organized into a 394-case purified subset and an 857-case extended subset, with case-level metadata covering label source, image-repair requirements, quality-control status, access conditions, checksums, and release boundaries. Compared with the original BraTS-GLI annotations, the resource substantially expands foreground supervision by incorporating healthy brain tissues and previously unlabeled coexisting abnormalities within a unified label space. A validation study using MedNeXt and T1/FLAIR inputs suggests that WMH-aware supervision preserves healthy-tissue segmentation performance across both in-domain GLI and external WMH datasets, while improving sensitivity to coexisting lesions relative to noisy-control training. The resource is intended for scientific research and supports joint anatomy-lesion supervision, label-noise analysis, and reproducible evaluation. Data are available at https://www.synapse.org/Synapse:syn75210889/wiki/, and code is available at https://github.com/xyx200/brats-gli-anatomy-lesion-code. The data resource DOI is https://doi.org/10.7303/SYN75210889.

Sources

Related papers