102 AI-Based Molecular Classification of Diffuse Gliomas using Rapid, Label-Free Optical Imaging.

Area of Science:

• Neuro-oncology
• Computational pathology
• Artificial intelligence in medicine

Background:

• Molecular classification is crucial for brain tumor management, but diagnostic testing is often delayed.
• Current molecular diagnostics can be time-consuming, impacting surgical and treatment decisions.
• Accurate and rapid molecular subtyping of diffuse gliomas is essential for personalized therapy.

Purpose of the Study:

• To develop and validate a deep learning-based system for rapid molecular classification of adult-type diffuse gliomas.
• To predict key molecular features (IDH, 1p19q-codeletion, ATRX loss) directly from histology during surgery.
• To provide a faster, scalable alternative to conventional molecular diagnostic methods.

Main Methods:

• Utilized stimulated Raman histology (SRH), a label-free optical imaging technique.
• Developed a multimodal deep neural network trained on SRH images and public genomic data (TCGA, CGGA).
• Prospectively validated the DeepGlioma system across multiple institutions.

Main Results:

• Achieved an average molecular genetic classification accuracy of 93.2% using the DeepGlioma system.
• Correctly identified diffuse glioma molecular subgroups with 91.5% accuracy within 2 minutes in the operating room.
• DeepGlioma demonstrated superior or comparable performance to conventional IDH1-R132H immunohistochemistry screening.

Conclusions:

• Artificial intelligence combined with optical histology offers a rapid and scalable solution for intraoperative molecular diagnosis of brain tumors.
• This approach can significantly expedite personalized treatment decisions for diffuse glioma patients.
• DeepGlioma has the potential to streamline clinical trial enrollment and improve patient outcomes.