OpenSRH

OpenSRH provides a dataset and computational framework that integrates stimulated Raman histology (SRH) imaging with deep learning to enable automated intraoperative multiclass histologic brain tumor classification and support real-time surgical decision-making.

Key Features:

• Dataset composition: Clinical SRH images from over 300 brain tumor patients and more than 1300 unique whole slide optical images, including raw and processed optical imaging data, pathologic annotations, and whole-slide tumor segmentations.
• Imaging modality: Stimulated Raman histology (SRH) whole-slide optical imaging as the primary data source for histologic assessment.
• Deep learning-based image interpretation: Automated image interpretation using deep learning for multiclass histologic brain tumor classification.
• Patch-based classification: Patch-level classification of SRH images using weak (patient-level) diagnostic labels.
• Contrastive representation learning: Support for patch-based contrastive representation learning and other advanced computer vision tasks.
• Multiclass histologic classification: Framework supports classification across multiple brain tumor diagnoses using SRH-derived features.
• Annotated whole-slide data: Detailed pathologic annotations and whole-slide tumor segmentations for training and validation of models.

Scientific Applications:

• Intraoperative diagnostic support: Automated SRH interpretation to inform surgical decision-making during brain tumor resection.
• Model development and validation: End-to-end development and validation of deep learning models using raw and processed SRH data, annotations, and segmentations.
• Representation learning and computer vision research: Benchmarking and development of patch-based contrastive representation learning and related CV methods on SRH data.

Methodology:

Patch-based classification of SRH images using weak (patient-level) diagnostic labels, multiclass histologic brain tumor classification, and patch-based contrastive representation learning.