Machine-learning-based tumor segmentation and classification using dynamic optical contrast imaging for thyroid...

Area of Science:

• Biomedical Optics
• Medical Imaging
• Machine Learning in Healthcare

Background:

• Thyroid cancer diagnosis is challenging due to overlapping benign/malignant nodule features.
• Fine-needle aspiration often yields indeterminate results and lacks spatial context, leading to unnecessary surgeries.
• Accurate, label-free, spatially resolved imaging is needed for intraoperative margin assessment.

Purpose of the Study:

• Develop and evaluate a machine learning-integrated Dynamic Optical Contrast Imaging (DOCI) framework.
• Classify thyroid tissue subtypes and segment cancerous regions from ex vivo hyperspectral sections.
• Assess the potential for real-time surgical application.

Main Methods:

• Imaged fresh ex vivo thyroid specimens using 23-channel DOCI.
• Employed pixel-level principal component analysis (PCA) and logistic regression for tissue classification.
• Utilized tumor-specific U-Net models for semantic segmentation and PCA-guided channel reduction.

Main Results:

• The first two PCA components explained over 70% of spectral variance, enabling well-separated tissue clusters.
• The PCA classifier achieved 92.3% validation accuracy and 100% test set accuracy.
• Reduced-channel models preserved high segmentation performance, improving follicular subtype accuracy.

Conclusions:

• DOCI integrated with interpretable machine learning enables accurate, label-free thyroid tissue differentiation and segmentation.
• Channel reduction demonstrates high performance with a compact spectral subset, supporting cost-efficient DOCI systems.
• This approach holds potential for future real-time intraoperative deployment in thyroid surgery.