Classifying Phonotrauma Severity from Vocal Fold Images with Soft Ordinal Regression

TL;DR

This work addresses automatic phonotrauma severity assessment from vocal fold images, treating severity as an ordinal variable and incorporating label uncertainty via soft labels. The authors propose soft ordinal regression (OR-Soft and CORAL-Soft), which extends standard ordinal regression losses to learn from annotator distributions. Through five-fold cross-validation on a dataset of 214 subjects, methods that leverage soft labels achieve calibration on par with or better than expert judgments, with CORN delivering the best predictive performance and OR-Soft providing the best balance of accuracy and uncertainty. The release of the dataset and code enables scalable, objective studies of phonotrauma and has potential to improve clinical understanding and patient care.

Abstract

Phonotrauma refers to vocal fold tissue damage resulting from exposure to forces during voicing. It occurs on a continuum from mild to severe, and treatment options can vary based on severity. Assessment of severity involves a clinician's expert judgment, which is costly and can vary widely in reliability. In this work, we present the first method for automatically classifying phonotrauma severity from vocal fold images. To account for the ordinal nature of the labels, we adopt a widely used ordinal regression framework. To account for label uncertainty, we propose a novel modification to ordinal regression loss functions that enables them to operate on soft labels reflecting annotator rating distributions. Our proposed soft ordinal regression method achieves predictive performance approaching that of clinical experts, while producing well-calibrated uncertainty estimates. By providing an automated tool for phonotrauma severity assessment, our work can enable large-scale studies of phonotrauma, ultimately leading to improved clinical understanding and patient care.

Figures (9)

• Figure 1: Images of vocal folds in the adducted (closed) position, showing varying levels of phonotrauma severity. Normal indicates healthy control.
• Figure 2: Overview of our proposed soft ordinal regression approach. (1) Ordinal Model Predictions: as in prior work niu2016ordinalcao2020rank, we train a model to perform $K-1$ tasks, where the $k$-th task is to predict whether the label exceeds rank $k$. (2) Soft Ordinal Labels: we form soft labels corresponding to the empirical probability that the label exceeds rank $k$. (3) Soft Ordinal Loss: we sum over task-specific weighted binary cross entropy loss terms; the weights correspond to soft label probabilities from part (2).
• Figure 3: Confusion matrices for CE-Soft and OR-Soft. Both methods have high accuracy in discriminating between normal and non-normal cases. OR-Soft makes fewer off-by-two errors than CE-Soft.
• Figure 4: Top: calibration curves for hard methods (blue) and their soft variants (orange). The orange curves are consistently closer to the gray line (perfect calibration) than the blue curves. Bottom: density of predicted probabilities for each method.
• Figure 5: Risk-coverage curves for the three methods with the lowest AURC. OR-Soft has the lowest risk for low to moderate coverage rates, whereas CORN has the lowest risk for high coverage rates.