Comparing classification methods for diffuse reflectance spectra to improve tissue specific laser surgery.

Area of Science:

• Oral and Maxillofacial Surgery
• Biomedical Engineering
• Machine Learning Applications

Background:

• Laser scalpels offer advantages in oral surgery but lack haptic feedback, posing risks to critical tissues like nerves.
• Real-time tissue classification using reflected light spectra can enhance surgical safety by warning or disabling lasers near sensitive structures.

Purpose of the Study:

• To analyze and compare various machine learning algorithms for classifying tissue types based on laser reflectance spectra.
• To identify the most suitable algorithm for real-time tissue discrimination in oral laser surgery.

Main Methods:

• Simulated larger datasets from a smaller experimental dataset using a multivariate Gaussian distribution.
• Trained and evaluated multiple machine learning algorithms on these datasets.
• Assessed performance using averaged confusion matrices and misclassification rates via boxplots, with validation on the original data.

Main Results:

• Most algorithms achieved a median misclassification rate below 0.25 on simulated data.
• Penalized Discriminant Analysis (PDA) demonstrated superior performance with a 0.00 misclassification rate on simulated data.
• PDA achieved an average misclassification rate of 0.02 in 10-fold cross-validation on the original dataset.

Conclusions:

• Penalized Discriminant Analysis is the most promising algorithm due to its ability to handle correlated spectral data.
• This study enhances real-time tissue discrimination accuracy.
• The findings represent a crucial advancement toward improving the safety of oral laser surgery.