Deep learning models for differentiating three sinonasal malignancies using multi-sequence MRI.

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Distinguishing between sinonasal squamous cell carcinoma (SCC), adenoid cystic carcinoma (ACC), and olfactory neuroblastoma (ONB) using MRI is challenging.
• Deep learning (DL) offers potential for improving diagnostic accuracy in complex oncological imaging.

Purpose of the Study:

• To develop and evaluate MRI-based deep learning (DL) models for differentiating sinonasal SCC, ACC, and ONB.
• To assess the impact of these DL models on the diagnostic performance of radiologists.

Main Methods:

• Retrospective analysis of 465 patients with sinonasal SCC, ACC, or ONB.
• Development of conventional MRI and DL models using T2WI, CE-T1WI, and ADC sequences.
• Evaluation of DL model performance using ResNet101, ResNet50, and DensNet121 architectures.
• Assessment of radiologist performance with and without AI assistance.

Main Results:

• The conventional MRI model achieved an AUC of 78.8%.
• The ResNet101 DL model demonstrated superior performance, with the mean fusion sequence yielding a macro-AUC of 0.892 and micro-AUC of 0.875.
• AI assistance significantly improved diagnostic accuracy, recall, precision, and F1-score for both senior and junior radiologists.

Conclusions:

• The ResNet101 DL model effectively differentiates sinonasal SCC, ACC, and ONB.
• AI-powered DL models enhance the diagnostic capabilities of radiologists, leading to improved accuracy in sinonasal tumor classification.