Laryngeal lesion classification stops relying on appearance alone: PADSNet adds vocal-fold vibration waveforms to a video model, lifting F1 about 0.20 over the strongest video baseline
Synopsis
In laryngoscopy, high-risk, low-risk and normal tissue often overlap in static appearance, while differences in vocal-fold symmetry, closure and periodicity may carry more discriminative information. PADSNet models video appearance features alongside glottal-distance waveforms extracted from vocal-fold segmentation masks, and uses nine biomechanical indices to gate and recalibrate the visual features. On a patient-level stratified three-class clinical video dataset of 397 patients, PADSNet reaches 0.8452 accuracy and 0.8089 F1, above the best compared video model, InternVideo2, at 0.6084.
Interpretation
Laryngeal lesion classification gains an evidence channel from vocal-fold vibration rather than static appearance alone. Earlier deep video models relied mainly on visual and spatiotemporal representations and did not explicitly include clinically interpretable biomechanical information such as bilateral symmetry, glottal closure, periodicity and vibration stability. On the three-class laryngoscopy video data from 397 patients, PADSNet reaches an F1 of 0.8089, while the highest among the compared video models is InternVideo2 at 0.6084; Video Swin Transformer and V-JEPA2 reach F1 scores of 0.5284 and 0.5993.
The two signals are split by frequency before interacting, and biomechanical indices then decide which features are amplified. The visual stream uses a Video Swin Transformer to extract spatiotemporal appearance features, while the biomechanical stream uses a bidirectional LSTM to model left and right glottal-distance sequences derived from segmentation masks. A wavelet-attention module applies a discrete wavelet transform to separate low-frequency trend and high-frequency detail in both streams, performs cross-attention within matching bands and reconstructs by inverse transform; a physics-aware gating module maps nine indices covering mobility, gap, symmetry and frequency-domain measures through a multilayer perceptron into channel gates that recalibrate features residually.
Perspective
The conclusions rest on laryngoscopy videos retrospectively collected from a single-center cohort of 397 patients at the Second Affiliated Hospital of Xi'an Jiaotong University, with labels confirmed by postoperative pathology and a patient-level stratified split of 80% training and 20% testing; the authors state that external multi-center validation is still needed to assess robustness across devices, acquisition protocols and patient populations.
The authors report that some misclassifications occur in borderline cases where adjacent categories have highly similar static appearance and, in some instances, highly similar glottal waveforms; physiological occlusion prevents reliable segmentation of glottal boundaries and sets the corresponding glottal distances to zero, weakening the biomechanical features; the current bivariate glottal-distance waveform and nine scalar priors mainly describe global bilateral motion and may not fully capture spatially localized mucosal-wave propagation.
