Enhancing Approaches to Detect Papilloma-Associated Hyperostosis Using a Few-Shot Transfer Learning Framework in Extremely Scarce Radiological Datasets

TL;DR

This paper presents a deep learning method to detect a rare radiological condition using very limited data, achieving better results than traditional approaches.

Contribution

A few-shot transfer learning framework using window shifting and pre-trained models for rare disease detection in scarce radiological data.

Findings

The proposed framework achieved a mean Dice Similarity Coefficient of 0.48 ± 0.06.

The baseline model failed to converge and had a clinically insignificant DSC of 0.09 ± 0.02.

The method effectively reduces instability and overfitting in extremely limited data scenarios.

Abstract

Background/Objectives: The application of deep learning models for rare diseases faces significant difficulties due to severe data scarcity. The detection of focal hyperostosis (PAH) is a crucial radiological sign for the surgical planning of sinonasal inverted papilloma, yet data is often limited. This study introduces and validates a robust methodological framework for building clinically meaningful deep learning models under extremely limited data conditions (n = 20). Methods: We propose a few-shot learning framework based on the nnU-Net architecture, which integrates an in-domain transfer learning strategy (fine-tuning a pre-trained skull segmentation model) to address data scarcity. To further enhance robustness, a specialized data augmentation technique called “window shifting” is introduced to simulate inter-scanner variability. The entire framework was evaluated using a rigorous 5-fold cross-validation strategy. Results: Our proposed framework achieved a stable mean Dice Similarity Coefficient (DSC) of 0.48 ± 0.06. This performance significantly outperformed a baseline model trained from scratch, which failed to converge and yielded a clinically insignificant mean DSC of 0.09 ± 0.02. Conclusions: The analysis demonstrates that this methodological approach effectively overcomes instability and overfitting, generating reproducible and valuable predictions suitable for rare data types where large-scale data collection is not feasible.

Figures