CKAD-Net: class-guided key-feature aggregation and adaptive decision network for vocal cord lesion prediction

Vocal cord lesions (VCL) that are not promptly diagnosed and treated may adversely affect patients' voice quality and speech communication in the short term, diminishing their quality of life. Long-term progression may lead to malignant tumors in the larynx, posing a threat to patients' health and life. To this end, this study proposes CKAD-Net, a vocal cord lesion prediction model based on category-guided key feature aggregation and adaptive decision-making. The category-guided key feature aggregation mechanism adaptively adjusts the attention distribution across different features for various lesion types, effectively enhancing category-relevant feature representations and improving the model's predictive performance. The adaptive decision mechanism employs a learnable adaptive weight factor to dynamically weight the prediction results between feature region representations and category-guided discriminative feature representations. This enables adaptive fusion of decision outcomes from both representations, generating more accurate decisions. Finally, this study validated the performance of the CKAD-Net model on coarse-grained and fine-grained vocal cord lesion prediction tasks using the hospital-private dataset VCLScopeData. The experimental results show that CKAD-Net achieves AUC and ACC values of (0.944 ± 0.002 [0.940, 0.948], 0.864 ± 0.008 [0.848, 0.880] and 0.929 ± 0.003 [0.923, 0.935], 0.730 ± 0.015 [0.700, 0.759] respectively, demonstrating superior performance compared to other models.