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

Thyroid cancer is the most common endocrine malignancy, and diagnosis is often challenging due to overlapping features between benign and malignant nodules. Fine-needle aspiration, the clinical gold standard, frequently yields indeterminate results and lacks spatial context, leading to unnecessary surgeries. Real-time margin assessment also remains limited. There is a need for accurate, label-free, spatially resolved imaging. Dynamic optical contrast imaging (DOCI), which measures autofluorescence lifetimes of endogenous fluorophores, offers a promising platform for intraoperative cancer detection.

We develop and evaluate a machine learning integrated DOCI framework to classify thyroid tissue subtypes and segment cancerous regions from ex vivo hyperspectral sections, with potential for real-time surgical use.

Fresh ex vivo thyroid specimens were imaged using a 23-channel DOCI acquisition. A pixel-level principal component analysis (PCA) and logistic regression classifier produced tissue probabilities, aggregated by a regional majority-vote gate to categorize specimens as normal, follicular, or papillary. Tumor-specific squeeze-and-excitation U-Net models were trained on voxel-only inputs for semantic segmentation. PCA-guided channel ablation identified a reduced spectral subset, and the pipeline was retrained using a compact 12-channel input.

The first two PCA components explained over 70% of spectral variance and yielded well-separated tissue clusters. The regional PCA classifier achieved 92.3% validation accuracy and 100% accuracy on the test set. Full 23-channel U-Net models delivered strong segmentation (papillary: Dice nonempty 0.829, balanced Dice 0.914; follicular: Dice nonempty 0.618, balanced Dice 0.809). Reduced-channel models preserved most performance and improved follicular segmentation (Dice nonempty 0.762), confirming spectral redundancy.

Integrating DOCI with interpretable machine learning enables accurate, label-free differentiation and segmentation of thyroid tissues. Channel reduction demonstrates that high performance is achievable with a compact spectral subset, supporting faster, more cost-efficient DOCI systems and future real-time intraoperative deployment.