Enhancing Approaches to Detect Papilloma-Associated Hyperostosis Using a Few-Shot Transfer Learning Framework in...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning models struggle with rare diseases due to data scarcity.
• Detecting focal hyperostosis (PAH) is vital for sinonasal inverted papilloma surgery but lacks sufficient data.

Purpose of the Study:

• To develop and validate a deep learning framework for rare disease analysis with limited data (n=20).
• To enable clinically meaningful predictions despite data scarcity.

Main Methods:

• A few-shot learning framework based on nnU-Net architecture.
• In-domain transfer learning by fine-tuning a pre-trained skull segmentation model.
• Data augmentation using 'window shifting' to simulate scanner variability.
• Evaluation via 5-fold cross-validation.

Main Results:

• The proposed framework achieved a mean Dice Similarity Coefficient (DSC) of 0.48 ± 0.06.
• This significantly outperformed a baseline model trained from scratch (mean DSC of 0.09 ± 0.02).

Conclusions:

• The methodological approach effectively overcomes instability and overfitting in rare data scenarios.
• The framework generates reproducible and valuable predictions where large-scale data collection is infeasible.