Combining nonlinear optical imaging and AI for sensory head and neck cancer surgery

Introduction

Conventional intraoperative diagnostic head and neck cancer methodologies, which rely extensively on tissue biopsies, are both time-consuming and invasive. The study explores the potential of multimodal nonlinear optical microscopy combined with deep learning for real-time cancer detection.

Material and method

A novel CARS/TPEF/SHG endomicroscope was analyzed on 15 head and neck cancer patients, with 20 frozen tissue samples. The images were processed through deep learning models designed for semantic segmentation, with the objective of classifying the tissue into multiple configurations. To account for inter-patient variability in tissue composition, the study employed a patch-based training approach and cross-validation. To assess the accuracy of the model in identifying tumor and non-tumor tissues, key metrics such as specificity, sensitivity, and F1-score were evaluated.

Results

The diagnostic accuracy using a six-class configuration resulted in a specificity of up to 98%. A three-class configuration, in which similar classes were consolidated into "Tissue to resect" and "Tissue to preserve," yielded balanced segmentation results, with sensitivity and specificity reaching 88% and 96%, respectively.. These findings highlight the potential clinical utility of the model in intraoperative settings, where real-time precision is of paramount importance.

Discussion

While the results are promising, they indicate the necessity for larger and more diverse datasets to enhance model robustness and generalizability for in vivo applications. This technique has the potential to transform surgical oncology through automated, image-guided interventions.

Financial support of Carl Zeiss Foundation project Sensorized surgery, no: P2022-06-004 is acknowledged.

Es liegt kein Interessenkonflikt vor.