Uncertainty estimation for margin detection in cancer surgery using mass spectrometry

Area of Science:

• Computational biology and bioinformatics
• Medical technology and instrumentation
• Machine learning in healthcare

Background:

• Rapid evaporative ionization mass spectrometry (REIMS) is an emerging technology for real-time tissue analysis during surgery.
• Accurate deep learning models are crucial for REIMS-based cancer margin detection, but face challenges from data noise and inter-patient variability.
• Existing models struggle to account for predictive uncertainty, limiting their clinical deployment.

Purpose of the Study:

• To integrate uncertainty estimation into deep learning models for REIMS-based cancer surgery margin detection.
• To improve the reliability and accuracy of tissue classification by factoring in predictive confidence.
• To address challenges of noise and patient variability in REIMS data acquisition.

Main Methods:

• Collected 693 spectra from 91 basal cell carcinoma patients using the iKnife device.
• Trained a Bayesian neural network and two baseline models for tissue classification and uncertainty estimation.
• Filtered training data by removing samples with high estimated uncertainty and compared model performance.

Main Results:

• Baseline models showed no performance improvement after data filtering due to unreliable uncertainty estimation.
• The proposed Bayesian model achieved statistically significant improvements: 75.2% balanced accuracy, 74.1% sensitivity, and 82.1% AUC after filtering uncertain samples.
• Further improvement in sensitivity to 88.2% was observed when uncertain samples were removed from the test data.

Conclusions:

• This study pioneers the application of uncertainty estimation for training and deploying deep learning models in REIMS for cancer surgery.
• Uncertainty estimation was used to quantify input noise and report predictive confidence, enhancing model robustness.
• Considering uncertainty in model development demonstrably improves the accuracy of REIMS-based margin detection systems.