AI-driven reflectance confocal microscopy for noninvasive diagnosis and accurate surgical margin assessment intra-operatively in oral cavity squamous cell carcinoma.

Abstract

6082 Background: Oral cavity squamous cell carcinoma (SCC) remains a common malignancy in the head and neck region, with challenges in tumor resection and recurrence prevention. Traditional methods like frozen-section analysis are limited by time delays, sampling errors, and tissue distortion. Reflectance Confocal Microscopy (RCM) provides a noninvasive alternative for real-time, high-resolution imaging, but interpreting RCM images accurately requires expert knowledge. Integrating artificial intelligence (AI) could improve the accuracy and reliability of RCM image interpretation for diagnosing SCC and assessing surgical margins. The integration of machine learning and artificial intelligence (AI) has the potential to enhance the accuracy and reliability of RCM image interpretation, providing a more efficient tool for diagnosing oral cavity SCC and assessing surgical margins in real-time during surgery. Methods: We developed an AI model using Google Cloud’s AutoML platform to classify RCM images for diagnosing oral cavity SCC and evaluating tumor margins. The dataset comprised 4,090 RCM images from 83 patients, including 1,998 images of benign tissue and 2,092 images of malignant tissue. The dataset was divided into training (80%), validation (10%), and test (10%) sets. A single-label classification approach was employed to differentiate benign and malignant tissue. Model performance was evaluated using sensitivity, specificity, accuracy, F1 score, and negative predictive value. Results: The AI model achieved an area under the curve (AUC) of 0.99, sensitivity of 98.09%, specificity of 95.00%, accuracy of 96.58%, and an F1 score of 96.70%. In comparison, expert human readers in our prior study achieved accuracies of 90.91% for normal tissue and 81.7% for tumor detection, highlighting the accuracy of the AI model's diagnostic performance. Conclusions: The combination of RCM imaging with AI-powered analysis provides an accurate, noninvasive method for real-time diagnosis and surgical margin assessment in oral cavity SCC. The AI-driven model has excellent sensitivity, specificity, and overall accuracy, offering a potentially efficient and reliable modality for the real-time evaluation of digital RCM images. This approach can reduce the time required for intraoperative margin assessment, minimize patient anesthesia time, and overcome challenges related to conventional histopathology, ultimately improving surgical outcomes in patients with oral cavity SCC.