AI‐Powered TAT‐Net Hyperspectral Imaging for Rapid and Accurate Differentiation of Sinonasal Inverted Papilloma and Malignancy
- 1Department of Otorhinolaryngology, National Health Commission Key Laboratory of Otorhinolaryngology Shandong Provincial Key Medical and Health Discipline of Qilu Hospital of Shandong University Jinan China
- 2Department of Biomedical Engineering Institute, School of Control Science and Engineering Shandong University Jinan China
Abstract
Background Sinonasal inverted papilloma (IP) and malignancies share overlapping clinical and endoscopic features, and conventional histopathology frequently requires time‐consuming adjunct immunohistochemistry. Hyperspectral imaging (HSI) can capture subtle spectral changes associated with malignant transformation that are not apparent on routine hematoxylin‐eosin (H&E) staining. Methods In this retrospective single‐center study, 176 H&E‐stained surgical specimens from 176 patients (89 with IP and 87 with sinonasal malignancies) were allocated at the patient/specimen level to training, validation, and test sets in a 6:2:2 ratio before any hyperspectral data processing. Hyperspectral calibration, selection of regions of interest (ROIs), spectral preprocessing, feature extraction, and quality control yielded 9,296 eligible ROI‐level hyperspectral samples: 5,578 in the training set, 1,859 in the validation set, and 1,859 in the test set. Joint spectral‐texture features were evaluated using several reference deep‐learning models and a Transformer‐Attention‐based Network (TAT‐Net). The primary outcome was the area under the receiver operating characteristic curve (AUC) in the test set. Results HSI revealed distinct spectral signatures between IP and sinonasal malignancies. All deep learning models demonstrated good discriminative performance (AUCs ≥ 0.85), with TAT‐Net achieving the best overall performance. On the independent test set, TAT‐Net achieved an AUC of 0.9656 (95% CI: 0.9571–0.9741) for distinguishing between IP and sinonasal malignancies. Conclusion AI‐powered TAT‐Net analysis of HSI enables rapid and accurate differentiation between IP and sinonasal malignancies on routine H&E sections. This approach has the potential to shorten diagnostic workflows, support early decision‐making, and help standardize the evaluation of sinonasal tumors.
Related
- Relative CT Number as a Quantitative Imaging Biomarker for Differentiating Sinonasal Polyp, Inverted Papilloma, and Squamous Cell Carcinoma: A Preliminary Retrospective Study
- Artificial Intelligence‐Optimized Image Biomarker Identification for Sinonasal Inverted Papilloma Diagnosis Using Image‐Enhanced Versus White‐Light Endoscopy
- Comprehensive multi-omic dissection and AI-prioritized target identification in inverted papilloma–associated sinonasal squamous cell carcinoma
- Factors Associated with Postoperative Recurrence in Inverted Sinonasal Papilloma: A Single-Center Cohort Study
- North American Delphi Consensus Study on Sinonasal Malignancy Survivorship Care
BibTeX
@article{zhao2026aipowered,
title = {{AI‐Powered TAT‐Net Hyperspectral Imaging for Rapid and Accurate Differentiation of Sinonasal Inverted Papilloma and Malignancy}},
author = {Hao‐Miao Zhao and Jia‐Chang Kong and Yun‐Ze Li and Wei Li and Xin Feng},
journal = {World Journal of Otorhinolaryngology-Head and Neck Surgery},
year = {2026},
month = oct,
doi = {10.1002/wjo2.70170},
url = {https://doi.org/10.1002/wjo2.70170},
note = {CC BY},
}
RIS
TY - JOUR
TI - AI‐Powered TAT‐Net Hyperspectral Imaging for Rapid and Accurate Differentiation of Sinonasal Inverted Papilloma and Malignancy
AU - Hao‐Miao Zhao
AU - Jia‐Chang Kong
AU - Yun‐Ze Li
AU - Wei Li
AU - Xin Feng
PY - 2026
DA - 2026/10/23
JO - World Journal of Otorhinolaryngology-Head and Neck Surgery
DO - 10.1002/wjo2.70170
UR - https://doi.org/10.1002/wjo2.70170
AB - Background Sinonasal inverted papilloma (IP) and malignancies share overlapping clinical and endoscopic features, and conventional histopathology frequently requires time‐consuming adjunct immunohistochemistry. Hyperspectral imaging (HSI) can capture subtle spectral changes associated with malignant transformation that are not apparent on routine hematoxylin‐eosin (H&E) staining. Methods In this retrospective single‐center study, 176 H&E‐stained surgical specimens from 176 patients (89 with IP and 87 with sinonasal malignancies) were allocated at the patient/specimen level to training, validation, and test sets in a 6:2:2 ratio before any hyperspectral data processing. Hyperspectral calibration, selection of regions of interest (ROIs), spectral preprocessing, feature extraction, and quality control yielded 9,296 eligible ROI‐level hyperspectral samples: 5,578 in the training set, 1,859 in the validation set, and 1,859 in the test set. Joint spectral‐texture features were evaluated using several reference deep‐learning models and a Transformer‐Attention‐based Network (TAT‐Net). The primary outcome was the area under the receiver operating characteristic curve (AUC) in the test set. Results HSI revealed distinct spectral signatures between IP and sinonasal malignancies. All deep learning models demonstrated good discriminative performance (AUCs ≥ 0.85), with TAT‐Net achieving the best overall performance. On the independent test set, TAT‐Net achieved an AUC of 0.9656 (95% CI: 0.9571–0.9741) for distinguishing between IP and sinonasal malignancies. Conclusion AI‐powered TAT‐Net analysis of HSI enables rapid and accurate differentiation between IP and sinonasal malignancies on routine H&E sections. This approach has the potential to shorten diagnostic workflows, support early decision‐making, and help standardize the evaluation of sinonasal tumors.
SN - 2589-1081
SN - 2095-8811
ER -