An interpretable multi-instance learning method for accurate differentiation of malignant and benign laryngeal lesions in laryngoscopy.

backgroundLaryngeal cancer is a significant global health issue with high mortality, and early diagnosis is critical for survival. Developing accurate diagnostic models for laryngoscopy can reduce potential repeated biopsies and lessen the patient burden, representing an urgent clinical need. However, existing artificial intelligence models often function as black boxes and are trained on single, pre-selected images, which does not reflect the clinical workflow where multiple images are assessed.

methodsWe conducted a retrospective study on 611 patients who underwent white light endoscopy (WLE) examinations. We developed an interpretable multi-instance learning network (IMIL-Net), which uses a patient-level of images. The model uses a Swin Transformer encoder and a gated attention pooling mechanism to produce a patient-level diagnosis and instance-level interpretability scores. We evaluated the model using a 5-fold cross-validation and compared it against four baseline models. We quantitatively validated the model's interpretability by comparing its attention scores against physician-annotated regions of interest (ROI) from 30 cases using the Mann-Whitney U test.

resultsIMIL-Net achieved the highest diagnostic performance, with a mean area under the curve (AUC) of 0.975 (95% CI 0.959-0.991), accuracy of 0.915 (95% CI 0.883-0.947), sensitivity of 0.876 (95% CI 0.803-0.949), and specificity of 0.945 (95% CI 0.910-0.980). This was superior to all baseline models, including non-MIL architectures (AUC 0.888-0.897) and a logistic regression model using only clinical data (AUC 0.898). The model's interpretability was quantitatively confirmed: physician-annotated ROI images (n=63) received a significantly higher median attention score (20.94%) compared to non-ROI images (n=78, 15.32%, [Formula: see text]).

conclusionThe proposed IMIL-Net provides a high-accuracy, interpretable, and clinically-aligned diagnostic solution. By analyzing complete patient examinations and providing a statistically validated decision-making process, this model represents a trustworthy tool for integration into the clinical otolaryngology workflow.