🧠 DeepSite-MRI: A Public Benchmark for Deep-Seated Intracranial Tumor Diagnosis from Biopsy-Confirmed MRI

NeurIPS 2026 · Evaluations and Datasets Track 📋 License: CC BY 4.0

📌 At a Glance

Property	Value
Task	4-class Preoperative MRI Classification
Modality	3D T1-Weighted Contrast-Enhanced (T1-CE) MRI
Cases	249 (biopsy-confirmed)
Classes	Glioma · PCNSL · Germinoma · Rare
Tumor Location	Thalamus · Basal Ganglia · Brainstem
Label Source	Stereotactic Biopsy (Histopathology)
Annotation	TSP Pipeline + Expert Review
Best Baseline	78.90% (ConvNeXt3D + Refined Prior)
Institution	Beijing Tiantan Hospital
Ethics	IRB Approved · Fully De-identified

---

🎯 Why DeepSite-MRI?

Deep-seated intracranial tumors—located in the thalamus, basal ganglia, and brainstem—are adjacent to critical neural nuclei controlling motor, language, and vital functions. This makes direct surgical resection extremely risky and often infeasible. Instead, clinicians must first perform stereotactic biopsy to determine the histopathological subtype, then administer targeted treatment accordingly:

• PCNSL is treated primarily with chemotherapy; surgical resection is harmful.
• Glioblastoma (GBM) requires surgery combined with radiochemotherapy.

However, biopsy itself carries risks of hemorrhage, infection, and sampling error due to suboptimal target selection. Predicting pathological subtypes from preoperative MRI alone would provide a vital reference for both clinical decision-making and biopsy target planning.

The fundamental obstacle: no public benchmark existed.

Existing datasets (BraTS, TCGA, etc.) cover hemispheric or superficial tumors. Deep-seated lesions differ substantially—they are typically smaller, exhibit blurrier boundaries, and sit within a more complex anatomical background. Crucially, radiological differentiation among Glioma, PCNSL, and Germinoma in the thalamus and basal ganglia is known to be unreliable without histological confirmation. No public dataset with biopsy-confirmed labels for this anatomical region has existed—until now.

---

🗂️ Dataset Overview

Cohort Demographics

Attribute	Value
Total cases	249
Male / Female	135 (54.2%) / 114 (45.8%)
Age range	4–86 years
Mean age ± SD	43.7 ± 23.2 years
Data source	Beijing Tiantan Hospital (>500 stereotactic biopsies/year)

Pathological Distribution

Class	Count	Proportion
Glioma	149	59.8%
PCNSL	72	28.9%
Germinoma	18	7.2%
Rare	10	4.0%

⚠️ The class distribution reflects real-world clinical incidence, not sampling bias. The imbalance is medically irreducible. Downstream experiments should account for class imbalance accordingly (e.g., stratified sampling).

Inclusion / Exclusion Criteria

Included if ALL of the following:

1. Imaging suggests a deep intracranial space-occupying lesion; multidisciplinary team (MDT) assessed open surgical resection as high-risk → stereotactic biopsy performed.
2. Postoperative histopathological report gives a definitive diagnosis of Glioma, PCNSL, Germinoma, or Rare.
3. A structurally complete, artifact-free preoperative T1-CE sequence is available.

Excluded if ANY of the following:

• Pathologically confirmed benign or non-neoplastic lesion (e.g., cavernous hemangioma, inflammatory pseudotumor).
• Missing core sequences or image registration failure.
• Patient/family refused data authorization.

---

⚙️ TSP: Two-Stage Scalable Spatial Prior Annotation Pipeline

Manual annotation of deep-seated tumors can take an experienced neuroradiologist up to 90 minutes per case due to blurry boundaries, complex anatomical backgrounds, and scarcity of reference cases. To make large-scale dataset construction feasible, we designed TSP, a fully automatic two-stage pipeline that:

• Requires no pre-trained in-domain classification model (unlike GradCAM/CAM-based methods)
• Requires no manual point or box prompts (unlike SAM/SAM2-based approaches)
• Reserves expert effort solely for final quality validation

TSP operates under a single-lesion assumption: each scan is presumed to contain one spatially contiguous dominant lesion. Cases with multifocal or diffusely enhancing patterns represent a known limitation.

---

📊 Benchmark Results

We benchmarked 10 representative 3D architectures under three prior conditions to characterize benchmark difficulty and research value:

Condition	Description
Without Prior	Pure T1-CE; no spatial guidance (pure vision baseline)
Coarse Prior	TSP Block 1 output only (VLM bounding boxes + smoothing)
Refined Prior	Full two-stage TSP output

These progressive conditions allow the main experiment to function inherently as an ablation study. The spatial prior is integrated via early fusion: the prior mask is concatenated with the T1-CE volume as an additional input channel, requiring no backbone modifications.

Full Results (Accuracy, Mean ± Std over 5-Fold Cross-Validation)

Backbone	Without Prior	Coarse Prior	Refined Prior	Gain
ConvNeXt3D	75.50 ± 3.90	77.10 ± 3.60	78.90 ± 3.20	+3.40%
UniFormer	74.50 ± 4.20	76.20 ± 3.80	78.00 ± 3.40	+3.50%
Swin3D	74.20 ± 4.10	75.80 ± 3.80	77.60 ± 3.50	+3.40%
TransUNet3D	72.00 ± 4.80	73.80 ± 4.40	75.50 ± 4.00	+3.50%
ResNet3D	67.27 ± 4.26	71.44 ± 6.14	74.61 ± 6.29	+7.34%
EfficientNet3D	71.00 ± 4.50	72.80 ± 4.10	74.50 ± 3.80	+3.50%
I3D	70.80 ± 6.00	72.50 ± 5.50	74.04 ± 5.12	+3.24%
DenseNet3D	61.92 ± 0.75	67.35 ± 2.34	73.61 ± 6.40	+11.69%
ViT3D	69.50 ± 5.20	71.50 ± 4.80	73.50 ± 4.20	+4.00%
AEFlow	68.00 ± 5.50	69.80 ± 5.00	71.50 ± 4.60	+3.50%

Average gain (Refined Prior vs. Without Prior): +4.76% across all 10 backbones.
Average additional gain (Coarse → Refined Prior): +2.44%, confirming the MLP refinement stage contributes independently beyond VLM coarse localization.

⚠️ Even the best model (ConvNeXt3D, 78.90%) falls well below the accuracy levels (typically >90%) reported for similar architectures on larger benchmarks such as BraTS2021. This is primarily due to the inherent visual ambiguity of deep-seated small lesions and severe class imbalance (Rare: 10 cases, 4.0%). Existing methods still have substantial room for improvement on this task.

---

📈 Data Efficiency Analysis

To assess whether existing architectures are data-efficient enough for this clinically challenging task, we evaluate four representative backbones under the Refined Prior setting across training set sizes of 30%, 50%, 70%, and 100%, with the test set held fixed.

Backbone	30% (~60)	50% (~100)	70% (~140)	100% (199)
ConvNeXt3D	67.15 ± 5.20	72.45 ± 4.40	76.10 ± 3.80	78.90 ± 3.20
Swin3D	61.80 ± 6.10	68.30 ± 5.00	74.20 ± 4.10	77.60 ± 3.50
ResNet3D	65.40 ± 8.15	69.75 ± 7.40	72.80 ± 6.85	74.61 ± 6.29
DenseNet3D	64.10 ± 8.50	68.45 ± 7.80	71.50 ± 7.10	73.61 ± 6.40

---

📁 Dataset Structure

DeepSite-MRI/
│
├── README.md
│
├── Dataset-DeepSite/                        ← nii images + labels
│   ├── sub-001/
│   │   └── *.nii.gz
│   ├── sub-002/
│   └── ...  (249 subjects total)
│
├── Coarse Prior/                    ← TSP Block 1 output (VLM-generated)
│   ├── sub-001/
│   └── ...
│
├── Refined Prior/                   ← TSP Block 2 output (full pipeline)
│   ├── sub-001/
│   └── ...
│
└── Code/                            ← Annotation & preprocessing scripts
    ├── TSP-pipeline.py
    ├── VLM-pipeline.py
    ├── VLMlabelbenchmark.py
    ├── Plabelbenchmark.py
    ├── dicomtonii.py
    ├── differentscaletrainse.py
    └── preprocess.py

---

📋 Evaluation Protocol

All baselines in the paper use:

• 5-fold stratified cross-validation (stratified by subtype to preserve class distribution per fold)
• Metrics: Accuracy, Precision, Recall, F1-score
• Prior integration: Early fusion — prior mask concatenated as an additional input channel to the T1-CE volume
• Optimizer: AdamW (lr = 5×10⁻⁵, weight decay = 1×10⁻⁴), cosine annealing with warm restarts (T₀ = 10 epochs), up to 250 epochs
• Regularization: Gradient clipping (max norm 2.0), dropout (p = 0.2), label smoothing (ε = 0.05)
• Hardware: NVIDIA RTX 4090

---

🔭 Dataset Value and Open Research Directions

DeepSite-MRI is designed as a shared infrastructure resource, not a single-task dataset. Its biopsy-confirmed labels, multi-level spatial annotations, full demographic metadata, and clinically grounded class distribution collectively support a spectrum of research directions.

Transferability of the Annotation Methodology

The TSP paradigm (zero-shot VLM coarse localization + per-subject MLP refinement + expert review) offers a reusable blueprint for data-scarce medical imaging domains. When conventional semi-automatic tools cannot be initialized due to lack of in-domain pretrained models, TSP demonstrates an alternative: coarse localization requires no training data (zero-shot), while refinement needs only a small set of pseudo-labels. This paradigm is transferable to other rare pathologies (e.g., rare dermatological or fundus lesions) or anatomical regions with prohibitively high annotation costs (e.g., spinal cord, skull base).

Learning Under Data-Limited Regimes

The clinically realistic and irreducible long-tail distribution (Rare: 4.0%, 10 instances) provides a medically grounded testbed for long-tail recognition, few-shot adaptation, and meta-learning—where misclassifying a rare subtype carries direct clinical consequences. The combination of biopsy-confirmed labels and multi-level spatial annotations further supports weakly supervised segmentation and data-efficient transfer learning.

Clinical Translational Directions

• Radiomic biomarker mining tailored to specific MRI phenotypic subtypes of deep-seated lesions
• AI-assisted biopsy target planning to improve sampling accuracy and reduce procedural risk
• Surgical risk stratification based on deep anatomical location and tumor characteristics

These directions collectively address the core clinical motivation: reducing reliance on invasive biopsies through reliable preoperative MRI analysis. DeepSite-MRI is the first public resource that grounds this clinical problem in histopathologically verified labels, enabling rigorous validation of clinical decision support systems.

---

⚠️ Limitations

1. Single-center constraint: DeepSite-MRI is currently single-center. Models may not generalize to cohorts with different scanning protocols or patient demographics. Multi-center expansion (target: 1,000+ cases) is actively underway.
2. Extreme class imbalance: The Rare category contains only 10 instances (4.0%), yielding ~2 samples per validation fold and introducing substantial evaluation instability (fold-wise std: 3%–6%).
3. Single-lesion assumption in TSP: Cases with multifocal or diffusely enhancing patterns fall outside the TSP modeling scope and represent a known failure mode.
4. Insufficient architectural data efficiency: Existing architectures are fragile in low-data regimes, with Transformer-based models particularly lacking inductive biases for small-sample scenarios.

The open-source TSP pipeline enables researchers worldwide to generate spatial annotations on their own biopsy cohorts at minimal cost, providing a path to break through single-institution data boundaries. This work provides a starting point, not an end point.