| Property | Value |
|---|---|
| Author | James Dolezal |
| License | GPL-3.0 |
| Framework | TF-Keras |
| Research Paper | Nature Publication |
What is thyroid-brs-v1?
The thyroid-brs-v1 is a specialized deep learning model designed to analyze H&E-stained pathologic images of thyroid neoplasms. It generates a BRAF-RAS Score (BRS) ranging from -1 (BRAF-like) to +1 (RAS-like), indicating the genetic expression similarity to BRAF-mutant and RAS-mutant tumors. Built on the Xception architecture, the model incorporates two dropout-enabled hidden layers for robust prediction.
Implementation Details
The model processes images at 299x299 pixels with 302x302 μm resolution. It utilizes a modified Reinhard normalizer for stain normalization and requires specific image standardization through TensorFlow. The architecture includes:
- Xception-based convolutional neural network backbone
- Two hidden layers (1024 width) with dropout (p=0.1)
- Adam optimizer with 0.0001 learning rate and 0.98 decay every 512 steps
- Training performed on 369 slides (116 BRAF-like, 271 RAS-like tumors)
Core Capabilities
- Accurate prediction of BRAF-RAS gene expression signatures
- Processing of H&E-stained pathology slides
- Research-focused analysis of thyroid neoplasms
- Integration with both TensorFlow and Slideflow frameworks
Frequently Asked Questions
Q: What makes this model unique?
This model uniquely combines deep learning with genetic expression analysis, providing a non-invasive method to predict BRAF-RAS scores from H&E images. It's specifically optimized for thyroid neoplasm analysis and includes robust image preprocessing techniques.
Q: What are the recommended use cases?
The model is strictly for research purposes, particularly in educational settings and pathology classification research. It should not be used for clinical decision-making or direct patient care without proper research protocol approval.