Predicting sinonasal inverted papilloma attachment using machine learning: Current lessons and future directions

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Inverted papilloma (IP) is often associated with hyperostosis on CT scans.
• Identifying IP tumor origin and attachment sites is crucial for surgical planning.
• Computed tomography (CT) is a key imaging modality for evaluating IP.

Purpose of the Study:

• To develop and evaluate a machine learning (ML) model for identifying IP attachment sites on CT images.
• To assess the performance of a deep learning segmentation algorithm (nnU-Net) in this task.
• To determine factors influencing the model's accuracy.

Main Methods:

• Retrospective review of 58 patients with IP treated at the institution.
• Manual segmentation of tumor attachment sites on CT scans by the operating surgeon.
• Application of a nnU-Net model for automated identification and segmentation of IP attachment sites.
• Evaluation using 5-fold cross-validation and Sørensen-Dice coefficient.

Main Results:

• The ML algorithm identified the IP attachment site in 55.2% of patients.
• Average Dice score was 0.34 (+/- 0.24), indicating moderate segmentation performance.
• The model performed better for maxillary sinus attachment sites (OR 4.6) and worse in revision surgeries (OR 0.11).

Conclusions:

• A state-of-the-art ML model demonstrated capability in identifying IP attachment sites.
• The model showed high fidelity in select cases, particularly for maxillary sinus origins.
• Larger and more diverse datasets are necessary for reliable clinical integration of this ML tool.