---

Model Card for SpineXtract

Model ID: DavidReineckeMD/SpineXtract-Transformer-MLP

Model Summary

SpineXtract is a transformer-based multilayer perceptron (Transformer-MLP) designed for the multiclass classification of spinal tumors from Stimulated Raman Histology (SRH) images.
The model was trained on multicenter data from the University Hospital Cologne (UKK, Germany) and evaluated on independent test cohorts from New York University (NYU), University of Michigan (UM), and Medical University of Vienna (MUV).
It provides near–real-time intraoperative tumor classification into four diagnostic entities: meningioma, schwannoma, metastasis, and ependymoma.

---

Model Details

Model Description

• Developed by: David Reinecke, MD
• Institution: Department of General Neurosurgery, University Hospital Cologne, Germany
• Funded by: German Spine Foundation and German Research Foundation
• Shared by: David Reinecke, MD
• Model type: ResNet50 + Transformer-MLP
• License: Academic research use only (non-commercial)
• Finetuned from: Custom self-supervised BYOL encoder (ResNet-50 backbone)

Model Sources

• Repository: https://github.com/DavidReineckeMD/SpineXtract
• Paper: (in Review, 2025)

---

Uses

Direct Use

This model can be used for research on intraoperative tissue classification using label-free SRH microscopy images. It outputs class probabilities.

Downstream Use

• Integration into experimental SRH analysis pipelines.
• Adaptation to other histologic classification tasks after fine-tuning.
• Evaluation of model explainability and uncertainty quantification.

---

Bias, Risks, and Limitations

• Model trained on SRH data from one academic center; limited representation of rare spinal tumors.
• Not validated on prospective streams.

Recommendations

Users should:

• Treat predictions as assistive suggestions only.
• Perform independent site-specific validation before deployment.

---

Training Details

Training Data

• Training site: University Hospital Cologne (UKK)
• Data type: Label-free SRH images (16-bit TIFF)
• Patch size: 300 × 300 px
• Training set: 1,258 SRH slides (>15 entities; 198 spinal slides used for fine-tuning)
• Validation set: 10% stratified split
• External test sites: NYU, UM, MUV (44 patients / 142 slides)
• Ground truth: FFPE H&E ± IHC/molecular analysis on separate tissue; blinded neuropathologist review.

Preprocessing

• Augmentations:
  random horizontal/vertical flip, sharpness (factor=2), Gaussian blur (σ=1), noise, autocontrast, solarize, erasing, affine transform (10°, translate [0.1, 0.3]), random resized crop (300 px).

Training Hyperparameters

Parameter	Value
Optimizer	AdamW (LR=0.001; β=[0.9,0.999]; weight decay=0.07)
Scheduler	Cosine annealing with 5% warm-up
Epochs	10 (early stopping with patience=5)
Precision	bf16-mixed
Loss	Cross-Entropy
Batch size	256
Seed	1000

Speeds, Sizes, Times

• Training hardware: 4× Nvidia A100 (80 GB)
• Framework: PyTorch 2.1.2 / Torchvision 0.10.10
• End-to-end SRH + inference: ≈ 5 min total

---

Evaluation

Testing Data

Independent multicenter test set (44 patients, 142 slides) from NYU, UM, MUV. No overlap with training data.

Factors

• Site variation
• Tumor subtype
• Patch count per patient

Metrics

Metric	Patient-level	Slide-level
Macro Balanced Accuracy	92.9% (95% CI 85.5–98.2)	92.2%
Macro AUROC	98.0% (95% CI 93.8–100)	96.5%
Sensitivity / Specificity	89.4% / 96.4%	89.6% / 95.2%
Calibration	Brier = 0.22; ECE ≤ 0.15
Decision Curve	Positive net benefit 0.1–0.9 thresholds

Results

• Best epoch: 10 (validation loss plateau after epoch 5).
• High-confidence threshold (τ): 0.77 for reliable patient-level decisions.

---

Model Examination

• Tokenization: 2048-D embeddings → 64 × 32-D tokens.
• Positional encoding: sinusoidal (sin/cos, base 10,000, max_seq_len=5000).
• Transformer layers: 3 (8 heads each, d_model=32, FFN=128).
• Dropout = 0.10; GELU activations; LayerNorm (ε=1e−5).

---

Environmental Impact

Parameter	Value
Hardware type	4× Nvidia A100 80 GB GPUs
Cloud provider	RAMSES HPC (University of Cologne)
Compute region	Germany

---

Technical Specifications

Model Architecture and Objective

Custom ResNet-50 backbone (BYOL pretraining) + Transformer-MLP head for multiclass classification (4 classes).
Objective: cross-entropy with calibrated outputs via Platt scaling.

Compute Infrastructure

• Environment: Containerized (CUDA 12.x + cuDNN 8.x)
• Reproducibility: identical software stack across all centers
• Metrics library: Scikit-learn 1.4.0

Hardware

• Training: 4 × Nvidia A100 (80 GB)
• Inference: Single GPU or CPU-supported
• Precision: bf16-mixed

Software

Python 3.10.13, PyTorch 2.1.2, Torchvision 0.10.10, Scikit-learn 1.4.0, Pydicom 2.4.3, tifffile 2023.9.26

---

Model Card Authors

David Reinecke, MD
University Hospital Cologne, Germany

---

Model Card Contact

📧 [email protected]