CT-based deep learning for identifying sinonasal inverted papilloma: a single-center retrospective study.

BACKGROUND

This study aimed to develop a CT-based deep learning model as a decision-support system for the classification of sinonasal inverted papilloma (IP), nasal polyps, and normal paranasal anatomy.

METHODS

A retrospective dataset of 871 patients who underwent paranasal CT imaging between 2020 and 2024 was analyzed. All diagnoses were confirmed histopathologically. Two representative CT slices per patient were selected at the patient level by an experienced radiologist blinded to final diagnosis. The dataset was split into 70% training and 30% testing sets, with 20% of the training data used for validation. No clinical variables were included; the model relied solely on imaging data. A ResNet50-based convolutional neural network and a Vision Transformer model were trained and compared.

RESULTS

The ResNet50 model achieved a test accuracy of 98.47%, outperforming the Vision Transformer model (79.65%). The model demonstrated high class-wise performance, particularly for inverted papilloma detection.

CONCLUSION

A CT-based deep learning model can serve as a decision-support tool for the classification of sinonasal inverted papilloma. Rather than replacing histopathology, this approach may assist clinical triage and complement conventional diagnostic workflows in a hybrid diagnostic strategy.