Deep learning-based optical coherence tomography image analysis of human brain cancer

Area of Science:

• Neuro-oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate intraoperative delineation of brain tumors, particularly in eloquent cortex areas, is crucial for effective cancer resection, patient survival, and quality of life.
• Current methods using optical attenuation values show promise but often overlook local textural information for robust predictions.

Purpose of the Study:

• To develop and validate a deep convolutional neural network (CNN) model for real-time intraoperative brain tumor delineation.
• To integrate optical coherence tomography (OCT) image data with co-occurrence matrix features for improved cancer and non-cancer tissue differentiation.

Main Methods:

• A deep ensemble CNN model was trained on 5,831 labeled OCT images from 7 patients.
• Co-occurrence matrix features were extracted from OCT images to synergize attenuation and texture characteristics.
• The model was evaluated on a holdout set of 4 patients without requiring beam profile normalization.

Main Results:

• The proposed CNN model achieved 93.31% sensitivity and 97.04% specificity in differentiating brain tissues.
• Segmentation maps generated by the model showed excellent agreement with established attenuation mapping methods.
• The approach demonstrated high accuracy without the need for reference phantom-based beam profile normalization.

Conclusions:

• The developed deep learning approach effectively delineates brain tumors and non-cancerous tissues in real-time using OCT imaging.
• This method synergizes optical attenuation and texture features for enhanced predictive accuracy.
• The findings have significant implications for improving surgical outcomes in neuro-oncology and facilitating clinical translation.