NeuroNasal: Advanced AI-Driven Self-Supervised Learning Approach for Enhanced Sinonasal Pathology Detection

• Published in: Sensors (Basel, Switzerland) Vol. 25; no. 8; p. 2369
• Main Authors: Atitallah, Nesrine, Ben Atitallah, Safa, Driss, Maha, Nahar, Khalid M. O.
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI AG 08.04.2025; MDPI
• Subjects: Accuracy; Algorithms; Artificial Intelligence; Automation; Blended learning; Classification; Comparative analysis; Computer-aided medical diagnosis; Disease; Endoscopy; Humans; Machine learning; Magnetic Resonance Imaging; Medical imaging; Neural networks; Paranasal Sinuses - diagnostic imaging; Paranasal Sinuses - pathology; Pathology; pathology detection; Patients; random forest; self-supervised learning; sinonasal; Sinuses; Supervised Machine Learning; Tomography, X-Ray Computed; Saudi Arabia; AI; sinonasal; self-supervised learning; random forest; pathology detection
• ISSN: 1424-8220; 1424-8220
• DOI: 10.3390/s25082369

Sinus diseases are inflammations or infections of the sinuses that significantly impact patient quality of life. They cause nasal congestion, facial pain, headaches, thick nasal discharge, and a reduced sense of smell. However, accurately diagnosing these diseases is challenging due to multiple factors, including inadequate patient adherence to pre-diagnostic protocols. By leveraging the latest developments in Artificial Intelligence (AI), there exists a substantial opportunity to improve the precision and effectiveness of classification of these diseases. In this study, we present a novel AI-based approach for sinonasal pathology detection, using Self-Supervised Learning (SSL) techniques and Random Forest (RF) algorithms. We have collected a new diagnostic imaging dataset, which is a major contribution to this study. The dataset contains 137 CT and MRI images meticulously labeled by expert radiologists, with two classes: healthy and unhealthy (sinus disease). This dataset is a useful asset for developing and evaluating AI-based classification techniques. In addition, our proposed approach employs the Deep InfoMax (DIM) model to extract meaningful global and local features from the imaging data with a self-supervised method. These features are then used as input for an RF classifier, which effectively distinguishes between healthy and sinonasal pathological cases. The combination of both DIM and RF provides efficient feature learning and powerful classification of sinus cases. Our preliminary results demonstrate the efficiency of the proposed approach, which achieves a mean classification accuracy of 92.62%. These findings highlight the potential of our AI-based approach in improving sinonasal pathology diagnosis.