Deep Learning-Based Intraoperative Dual-tracer Video Analysis of Sentinel Lymph Node Mapping for Metastasis Prediction in cN0 Papillary Thyroid Carcinoma

BACKGROUND AND RATIONALE:

Papillary thyroid carcinoma (PTC) is one of the fastest-growing cancers worldwide. A major challenge in treating PTC is that 30% to 80% of patients who appear to have no lymph node involvement before surgery (clinically node-negative, or cN0) actually have hidden (occult) cancer spread to their lymph nodes. Current imaging methods like ultrasound often miss these small areas of cancer spread.

This creates a difficult decision for surgeons: removing too many lymph nodes increases the risk of complications such as damage to the parathyroid glands (which control calcium levels) and the nerves that control the voice. However, removing too few lymph nodes may leave cancer behind, which can lead to recurrence.

Sentinel lymph node (SLN) mapping is a technique that identifies the first lymph nodes that drain from a tumor. The idea is that if cancer spreads through the lymphatic system, it will reach these sentinel nodes first. However, current single-tracer methods for SLN mapping in thyroid cancer have limitations and variable results.

This study uses a dual-tracer approach that combines two different dyes:

1. Carbon nanoparticles (CNs): These provide long-lasting black staining that helps surgeons see lymph nodes clearly
2. Indocyanine green (ICG): This dye glows under near-infrared light, allowing real-time visualization of lymphatic flow

By combining these two tracers, surgeons can see both the structure of lymph nodes and how lymphatic fluid flows through them over time.

STUDY DESIGN:

This is a prospective, single-center, observational cohort study. The study does not change the surgical treatment that participants receive. All participants undergo standard thyroid cancer surgery with lymph node removal as determined by their surgical team.

STUDY PROCEDURES:

1. Pre-operative Assessment: All participants undergo standard pre-operative evaluation including:

• Physical examination
• Thyroid ultrasound with detailed lymph node assessment
• Fine-needle aspiration biopsy to confirm PTC diagnosis
• Genetic testing for common mutations (such as BRAF)
• Standard blood tests

1. Surgical Procedure: During surgery, participants receive the dual-tracer injection under ultrasound guidance. The injection is given at multiple points around the thyroid tumor. The specific preparation is:

• 0.1 ml of ICG solution (concentration: 2.5 mg/ml)
• 0.1 ml of carbon nanoparticle suspension (concentration: 50 mg/ml) These are mixed together and injected using precise, multi-point technique.

1. Video Recording: A near-infrared fluorescence imaging system records the entire process of lymph node visualization. The recording captures:

• How the dyes spread through the lymphatic channels
• When each lymph node first becomes visible
• How the fluorescence signal changes over time
• The pattern of lymphatic drainage Videos are recorded at high resolution (1920 × 1080 pixels) at approximately 30 frames per second. A standardized 3-minute segment is extracted from each video for analysis, providing 150 frames per patient.

1. Surgical Decisions: The sentinel lymph node (the first node that lights up) is removed and sent for immediate frozen section analysis. Based on standard criteria, surgeons decide whether to perform:

• Ipsilateral central lymph node dissection (removing lymph nodes on the same side as the tumor)
• Lateral lymph node dissection (if certain criteria are met)
• Contralateral central dissection (removing lymph nodes on the opposite side) These decisions follow the standard surgical protocol at our institution and are not influenced by the deep learning predictions.

1. Pathological Examination:

All removed lymph nodes are examined by pathologists to determine:

• Whether the sentinel lymph node contains cancer (SLNM status)
• Whether second-echelon lymph nodes contain cancer (SeLNM)
• Whether non-sentinel lymph nodes contain cancer (NsLNM)
• The number and location of all positive lymph nodes

DATA COLLECTION AND ANALYSIS:

Clinical Data (32 variables):

• Demographics: age, sex, body mass index
• Ultrasound features: tumor size, location, margins, calcification patterns, aspect ratio, TI-RADS classification
• Pathological features: multifocality, extrathyroidal extension, capsular invasion
• Genetic data: BRAF mutation status and other relevant mutations

Video Analysis:

Two experienced surgeons (each with more than 10 years of experience) manually identify and outline the regions of interest (the sentinel lymph nodes) in each video frame. This creates 19,650 mask images across all participants.

Feature Extraction:

The deep learning system extracts multiple types of features:

Spatial Features (2,048 dimensions):

• Image features from a pre-trained neural network (EfficientNet-B5)
• Grayscale characteristics
• Shape and morphological features
• Hu moment descriptors (mathematical descriptions of shape)

Temporal Features (20 dimensions):

• Frame-to-frame differences showing how the image changes
• Optical flow measurements showing movement patterns
• Fluorescence intensity changes over time
• Flow velocity measurements at specific time points

DEEP LEARNING MODELS:

Nine different deep learning architectures are developed and compared:

1. Convolutional Neural Network (CNN): Extracts local spatial features from images
2. Long Short-Term Memory (LSTM): Captures patterns in time-series data
3. CNN + LSTM: Combines spatial and temporal feature extraction
4. CNN + LSTM + Attention: Adds attention mechanism to focus on important time points
5. Transformer: Uses self-attention to capture global patterns in sequential data
6. Crossformer: A specialized architecture for time-series analysis
7. 3D-CNN: Processes video data as three-dimensional volumes
8. LSTM + Transformer: Hybrid model combining LSTM and Transformer strengths
9. LSTM + Crossformer: Hybrid model combining LSTM and Crossformer

All models use:

• Binary cross-entropy loss function
• Adam optimizer for training
• Data augmentation (rotation, scaling, noise injection) to prevent overfitting
• Weighted loss functions to handle class imbalance
• Dropout regularization

MODEL EVALUATION:

Models are evaluated using 10-fold stratified cross-validation, ensuring balanced distribution of outcomes in training and testing sets. Performance metrics include:

• Area Under the ROC Curve (AUC): Measures overall discrimination ability
• Accuracy: Percentage of correct predictions
• Sensitivity: Ability to correctly identify patients with metastasis
• Specificity: Ability to correctly identify patients without metastasis
• Positive Predictive Value: Probability that a positive prediction is correct
• Negative Predictive Value: Probability that a negative prediction is correct
• F1 Score: Balance between precision and recall
• Brier Score: Calibration of predicted probabilities

Additional analyses include:

• Decision curve analysis to assess clinical utility
• Calibration curves to check prediction reliability
• Learning curves to assess overfitting
• DeLong test for statistical comparison between models
• Probability-based model ranking approach (PMRA)

MODEL INTERPRETABILITY:

To understand how the model makes predictions, we use SHapley Additive exPlanations (SHAP) analysis. This technique:

• Identifies which features contribute most to predictions
• Shows how each feature affects individual predictions
• Reveals whether the model relies on clinically meaningful factors
• Provides transparency into the "black box" of deep learning

OUTCOMES:

Primary Outcomes:

1. Second-echelon lymph node metastasis (SeLNM): Cancer spread to lymph nodes beyond the sentinel node in the drainage pathway
2. Non-sentinel lymph node metastasis (NsLNM): Cancer spread to any lymph node other than the sentinel node

Both outcomes are determined by final pathological examination of surgically removed tissue (the gold standard).

Secondary Outcomes:

• Model performance metrics (AUC, sensitivity, specificity, etc.)
• Feature importance rankings from SHAP analysis
• Comparison of model architectures

STATISTICAL CONSIDERATIONS:

Sample Size:

Based on power calculations assuming:

• Alpha = 0.05 (significance level)
• Power = 0.80
• Expected sensitivity difference of 20% between methods
• Expected prevalence of lymph node metastasis of 50%

A minimum of 335 participants was calculated. Due to strict inclusion criteria and video quality requirements, 131 participants with complete, high-quality data were included in the final analysis.

Statistical Methods:

• Descriptive statistics for baseline characteristics
• Chi-square or Fisher exact test for categorical variables
• Mann-Whitney U test for continuous variables
• DeLong test for comparing AUC values between models
• 95% confidence intervals for all performance metrics

FOLLOW-UP:

While the primary analysis focuses on intraoperative prediction, participants are followed according to standard clinical care protocols. Long-term outcomes including recurrence-free survival may be analyzed in future studies.

ETHICAL CONSIDERATIONS:

This study was approved by the Ethics Committee of the First Affiliated Hospital of Chongqing Medical University (Approval No. 2023-322). All participants provided written informed consent before enrollment.

The study poses minimal additional risk to participants because:

• The dual-tracer injection uses agents already approved for clinical use
• The video recording is non-invasive and does not affect surgical decisions
• All surgical decisions are made according to standard protocols
• The deep learning analysis is performed retrospectively and does not influence treatment

POTENTIAL IMPACT:

If successful, this approach could:

1. Help surgeons make more informed decisions about the extent of lymph node removal
2. Reduce unnecessary extensive surgery in low-risk patients
3. Ensure adequate surgery in high-risk patients who might otherwise be undertreated
4. Provide real-time, objective decision support during surgery
5. Standardize the interpretation of lymphatic mapping across different surgeons and centers
6. Serve as a training tool for less experienced surgeons

LIMITATIONS:

• Single-center study, which may limit generalizability
• Relatively small sample size due to strict quality requirements
• Manual annotation of regions of interest may introduce variability
• Deep learning models require validation in external cohorts
• Real-time implementation during surgery requires further development

FUTURE DIRECTIONS:

• External validation in multiple centers
• Development of real-time prediction software for intraoperative use
• Integration with other imaging modalities
• Prospective interventional trials to assess clinical benefit of AI-guided surgery