OpenSRH: optimizing brain tumor surgery using intraoperative stimulated Raman histology

Area of Science:

• Neuro-oncology
• Computational Pathology
• Medical Imaging

Background:

• Standard intraoperative brain tumor diagnosis is time-consuming and resource-intensive, limiting surgical options.
• Rapid, accurate diagnosis is crucial for effective brain tumor surgery and patient care.
• There is a need for advanced tools to support real-time surgical decision-making.

Purpose of the Study:

• To introduce OpenSRH, the first public dataset of clinical stimulated Raman histology (SRH) images for brain tumor analysis.
• To enable the development and validation of deep learning models for intraoperative brain tumor diagnosis.
• To facilitate real-time surgical decision support and improve access to optimal cancer surgery.

Main Methods:

• Development of a workflow combining stimulated Raman histology (SRH) with deep learning for automated image interpretation.
• Creation of the OpenSRH dataset, including 1300+ whole slide images from 300+ brain tumor patients with detailed annotations.
• Implementation of a framework for patch-based whole slide SRH classification using weak diagnostic labels.

Main Results:

• The OpenSRH dataset provides comprehensive data for end-to-end model development, covering common brain tumor types.
• Benchmarking of multiclass histologic brain tumor classification and contrastive representation learning on the dataset.
• Demonstration of a framework for automated interpretation of SRH images for diagnostic support.

Conclusions:

• OpenSRH dataset and framework support the clinical translation of rapid optical imaging and machine learning in neuro-oncology.
• This approach can enhance the safety, efficacy, and accessibility of brain tumor surgery.
• Facilitates advancements in precision medicine for cancer treatment through AI-driven diagnostics.