Author(s)
Sean Mckee, MD
Luca Giancardo, PhD
William Yao, MD, FARS
Brady Anderson, MD
Jumah Ahmad
Andy Chua, MD
Chinmay Mokashi
Samia Islam
Salman Hasan, MD
Martin Citardi, MD, FARS
Amber Luong, MD, FARS
Affiliation(s)
University of Texas Health Science Center in Houston;
Abstract:
Background: Surgically addressing the origin of inverted papilloma (IP) is key to its complete resection, which is critical in preventing tumor recurrence. Areas of hyperostosis on computed tomography (CT) scans provide an indication of the IP’s origin, but is sometimes hard to discern. Herein, we developed a machine learning model to analyze CT images and assist in identifying IP attachment sites.
Methods: A retrospective review of patients treated for IP at our institution between 2004 and 2021 was conducted. The IP tumor attachment site was manually segmented on CT by the operating surgeon. We used a nnU-Net model, a deep learning-based segmentation algorithm that automatically configures image preprocessing, network architecture, training, and post-processing to identify the IP attachment site. The model was trained and evaluated using a 5-fold cross validation, where each iteration split the data into train/validation/test to avoid chances of overfitting. The Sørensen–Dice coefficient (Dice) was used to evaluate the segmentation performance of the nnU-Net model.
Results: A total of 68 subjects were included in the nnU-Net model. Thirty-nine (50%) subjects passed the initial preprocessing phase. The tumor attachment site was correctly identified by the nnU-Net model output in 22 (56%) subjects with an average Dice score of 0.345 and standard deviation of 0.237.
Conclusions: The nnU-Net deep learning model was able to successfully identify the sino-nasal attachment site of IP with a high degree of fidelity in selected cases. Machine learning shows early promise as a novel technology for identifying IP tumor origin. Future work is needed to fine-tune the model for greater accuracy.