Optimization research on a laryngoscopic vocal fold lesion segmentation model integrating multi-scale edge awareness and transformer
Laryngoscopy is the primary clinical modality for evaluating vocal fold lesions; however, its interpretation is highly dependent on the subjective experience of the physician, making diagnostic consistency difficult to guarantee. To address the insufficient boundary delineation of existing deep learning methods for laryngoscopic image segmentation, this study proposes MEA-TransNet, an optimized vocal fold lesion segmentation model integrating multi-scale edge awareness with a Transformer. The task is formulated as binary pixel-level segmentation across five clinical categories, with normal vocal fold images (empty reference masks) retained as negative controls for assessing false-positive segmentation. The model adopts ResNet-34 as the backbone with a multi-scale edge-aware module (MEAM) that extracts lesion boundary features through the synergistic interaction of Sobel gradient priors and learnable depthwise separable convolutions with multiple dilation rates. A lightweight windowed self-attention Transformer encoder is introduced at the deep layers for global contextual modeling at controllable computational cost. Images from 862 patients, with leukoplakia and early glottic squamous cell carcinoma (Tis to T2N0M0) histopathologically confirmed, were split at the patient level with category stratification; 95% confidence intervals (CI) were obtained by patient-level bootstrap. On a primary dataset of 4,216 laryngoscopic images covering normal vocal folds and four lesion types, MEA-TransNet achieves a Dice Similarity Coefficient of 89.14% (95% CI 88.03 to 90.22) and an Intersection over Union of 81.93%, representing improvements of 2.21 and 3.19 percentage points over the strongest baseline nnU-Net, and 3.02 and 4.07 percentage points over DA-TransUNet, respectively. Patient-level five-fold cross-validation yields a mean DSC of 88.98% ± 0.32 percentage points. Zero-shot evaluation on an independent external cohort of 487 images from 124 patients at a second tertiary hospital using a different laryngoscope manufacturer gives a DSC of 86.47% (95% CI 84.78 to 88.08). On the boundary-ambiguous categories of leukoplakia and early glottic carcinoma, MEA-TransNet outperforms nnU-Net by 3.12 and 3.17 percentage points in DSC ( P = 0.004 and 0.008, Holm-corrected), and on the external early glottic carcinoma subgroup (21 patients) the patient-level DSC is 82.4% (95% CI 78.1 to 86.5). MEA-TransNet improved the delineation of vocal fold lesions, with the largest gains on the boundary-ambiguous categories of leukoplakia and early glottic carcinoma, and maintained its performance on an external cohort acquired with a different laryngoscope. The findings suggest a promising assistive tool for lesion delineation; prospective reader studies, video-based validation, and larger multicenter cohorts are required before clinical use.
Authors
- Xiaoyi Yang
Institutions
- Pingdingshan University (CN)
Publication Details
- Journal
- European journal of medical research
- Published
- 2026-11-20
- DOI
- https://doi.org/10.1186/s40001-026-05487-2
- Primary Topic
- Medical Image Segmentation Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00