Generative AI sharpens laryngeal lesion classification from radiomic features

Researchers have developed a generative artificial intelligence framework that improves the classification of laryngeal lesions without creating a single synthetic medical image. The approach, described in an open-access study in BMC Medical Imaging, trains generative models directly in the radiomic feature space, the high-dimensional space of quantitative descriptors extracted from medical images, rather than fabricating new pictures. In an evaluation on an independent, unseen test set, the framework’s best-performing configuration, Radiomics-GAN, achieved a median F1-score of 0.851, outperforming a baseline model balanced with SMOTE but without generative augmentation, which reached 0.806.

The study addresses a persistent bottleneck in computer-assisted diagnosis of laryngeal cancer. Narrow-band imaging, or NBI, is an endoscopic technique that enhances mucosal and vascular contrast and has become an important modality for the early detection of laryngeal cancer. Radiomics can extract high-dimensional quantitative features from NBI images, feeding machine-learning classifiers that assist diagnosis. However, the availability and heterogeneity of annotated datasets remain limited, which constrains both the development of machine-learning models and their generalizability to new patients and imaging conditions.

Rather than generating synthetic NBI images, the proposed framework trains generative models to learn realistic distributions of radiomic features and to produce valid synthetic radiomic representations. By enriching the radiomic feature manifold in this way, the authors report, the method effectively mitigates sample scarcity and improves classifier performance. The framework was evaluated with multiple machine-learning classifiers on a laryngeal cancer diagnosis task, and generative augmentation consistently improved classification performance across all classifiers tested. Among the generative models evaluated, Generative Adversarial Networks delivered the most consistent gains.

The dataset comprised 11,144 NBI images from 210 adult patients, divided into a development set of 180 patients and an independent internal held-out test set of 30 patients. Performance was assessed on the unseen test set using accuracy, precision, recall and F1-score. The reported advantage of Radiomics-GAN over the SMOTE-balanced baseline indicates that feature-space augmentation can outperform conventional resampling approaches for this task, although the held-out test cohort was small and the evaluation remained internal.

The authors conclude that the framework supports the development of more robust AI-assisted tools for laryngeal cancer diagnosis. Future work, they write, will focus on external validation and on conditional generative models tailored to lesion subtypes and to variability in imaging acquisition, directions that will show how the reported gains extend beyond this internal evaluation.

Subject of Research: Generative AI augmentation of radiomic features for laryngeal lesion classification on narrow-band imaging

Article Title: Radiomics-GenAI: a novel model for laryngeal lesion classification

Article References: Wang, H., Xie, Z., & Mainardi, L. (2026). Radiomics-GenAI: a novel model for laryngeal lesion classification. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02963-3

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02963-3

Keywords: radiomics, generative AI, GAN, laryngeal lesions, narrow-band imaging, medical imaging, machine learning, feature augmentation, computer-assisted diagnosis, laryngeal cancer, Radiomics-GenAI, novel