Validation and correlation with clinical data of a newly developed computer aided diagnostic system for the classification of paranasal anomalies in the maxillary sinus from MRI images

Introduction Large scale population studies are used to analyse the rate of finding sinus opacities in cranial MRIs (cMRI). Artificial Intelligence support systems can automate the detection of sinus opacities and reduce the workload of clinicians. We developed and evaluated a Computer Aided Diagnostics system based on a 3D Convolutional Neural Network (3D CNN) that automatically extracts and classifies maxillary sinus (MS) from cMRI.

Methods As part of the Hamburg City Health Study, cMRIs of 2619 participants (45-74 years) were recorded for neuroradiological assessment. 1069 participants MS were manually diagnosed for incidental findings by 2 ENT specialists and a ENT-specialized radiologist. The labelled dataset was used to develop and train a 3D CNN that extracts MS from cMRI and classified MS with opacifications (polyps, cyst, mucosal thickening) from MS without opacifications. Association of the two groups with multiple clinical data was tested.

Discussion We extracted 30 MS volumes from each participants MRI. The 3D CNN dataset included 19215 (59.91%) MS without opacifications, 4815 (15.01%) with mucosal wall inflammations (>2mm), 6315 (19.69%) with polyps, 1185 (3,69%) with cysts, and 540 (1.68%) with polyps/cysts encompassing the entire MS. The evaluation metrics (AUROC: 0.95, sensitivity: 0.85, specificity: 0.90) demonstrated the effectiveness of our approach. Statistically significant associations between the two groups were observed regarding alcohol consumption, BMI, asthma, hay fever and sex.

Conclusions Our 3D CNN showed the ability to classify MS with and without opacifications and automatically diagnose incidental findings, which can enhance the efficiency of uncovering correlations with clinical data in the context of population studies.

Publication History

Article published online:
19 April 2024