Towards Precision Medicine in Sinonasal Tumors: Low-Dimensional Radiomic Signature Extraction from MRI.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Sinonasal tumors represent a rare subset of head and neck neoplasms (3-5%).
• Existing machine learning (ML) and radiomics models for tumor classification often lack detailed morphological and textural characterization.
• Advanced imaging analysis is needed for precise sinonasal tumor diagnosis.

Purpose of the Study:

• To evaluate the effectiveness of machine learning (ML) and radiomics in classifying sinonasal tumors.
• To assess the added value of integrating clinical variables with radiomic features from MRI.
• To explore the potential of the DNetPRO algorithm for sinonasal tumor signature extraction.

Main Methods:

• Analysis of multi-center MRI data from 145 patients (76 malignant, 69 benign).
• Extraction of radiomic features from T1-weighted (T1-w) and T2-weighted (T2-w) MRI sequences.
• Development of an ML pipeline to evaluate radiomic features and their integration with clinical data, utilizing the DNetPRO algorithm.

Main Results:

• ML classification combining radiomic and clinical data achieved a median Matthews Correlation Coefficient (MCC) of 0.60 ± 0.07.
• The DNetPRO algorithm yielded the highest MCC of 0.73 using both T1-w and T2-w images.
• Clinical factors like symptoms and tumor size were significant, complemented by radiomic insights into texture and gray-level distribution.

Conclusions:

• ML-based radiomics holds potential for sinonasal tumor classification but faces challenges in clinical adoption due to data variability.
• Standardization and interpretability are critical for the reliable clinical implementation of radiomics models.
• The DNetPRO approach demonstrates the clinical relevance of integrating radiomic and clinical data for improved sinonasal tumor diagnosis.