SIGNAL: A Scalable, Real-World Model for Rapid Intraoperative Molecular Classification of Gliomas Using Stimulated...

Area of Science:

• Neuro-oncology
• Computational pathology
• Medical imaging analysis

Background:

• Previous machine learning models for intraoperative glioma molecular prediction using stimulated Raman histology (SRH), like DeepGlioma, showed high accuracy on curated datasets but struggled with real-world clinical application due to slow processing and lack of image quality control.
• DeepGlioma's large parameter count (162M) limited its intraoperative speed on existing SRH hardware, and its validation on curated images did not reflect the complexities of uncurated clinical data.

Purpose of the Study:

• To develop a novel, lightweight machine learning model for intraoperative molecular classification of diffuse gliomas using SRH.
• To improve upon the performance and speed of existing models by incorporating an attention-based mechanism for image quality filtering.
• To enable robust and generalizable intraoperative glioma classification on uncurated clinical datasets.

Main Methods:

• Developed SIGNAL (SRH-Informed Glioma classificatioN with Attention Learning), a 27M parameter model, using 1.56 million SRH images from 967 adult diffuse glioma patients.
• Employed a ResNet50 backbone pre-trained with hierarchical contrastive loss, followed by a multi-head multi-layer perceptron (MLP).
• Implemented a patch-based attention threshold (0.6) and a final MLP for predicting glioma subtypes and molecular status (IDH mutation, 1p19q codeletion, ATRX loss).

Main Results:

• SIGNAL achieved 90.10% overall accuracy in glioma subtype classification, significantly outperforming DeepGlioma (72.59%) and operating faster (16.0 vs. 6.7 patches/s).
• SIGNAL demonstrated superior accuracy in molecular classification: IDH mutation (93.51% vs. 79.22%), 1p19q codeletion (93.51% vs. 88.31%), and ATRX loss (89.61% vs. 83.98%).
• The attention mechanism effectively filtered low-quality patches (e.g., blood, acellular regions), showing strong correlation with cellularity and inverse correlation with blood coverage, while maintaining accuracy in tumor margins.

Conclusions:

• SIGNAL is a highly accurate and efficient lightweight model for intraoperative molecular classification of diffuse gliomas via SRH.
• The integrated attention mechanism enhances robustness by filtering diagnostically uninformative image regions, crucial for reliable intraoperative decision-making.
• SIGNAL offers a significant advancement for intraoperative glioma diagnosis, providing faster, more accurate, and interpretable results compared to previous state-of-the-art models.