Automated Classification of Maxillary Sinus Anomalies Using Self-Supervised Learning
Overview
This study introduces a novel self-supervised learning (SSL) approach to classify maxillary sinus (MS) anomalies from 3D MRI data. By leveraging a 3D convolutional autoencoder to generate pseudo segmentation masks of anomalies, the method improves downstream classification performance distinguishing normal from anomalous MS.
Background
Paranasal sinuses, including the maxillary sinuses, exhibit significant anatomical variability and are prone to pathologies such as retention cysts, polyps, and mucosal thickening. Accurate diagnosis using 3D imaging modalities like MRI is critical but challenging due to incidental findings and variability in sinus appearance. Prior deep learning approaches have relied on supervised learning, which requires extensive labeled datasets that are difficult to obtain. Self-supervised learning offers a promising alternative by utilizing unlabelled data to improve feature representation for anomaly classification.
Data Highlights
Dataset
Number of Patients
Normal MS
Anomalous MS
Labelled Dataset
1067
489
578
Unlabelled Dataset
1559
Not specified
Key Findings
The proposed SSL task uses residual volumes generated by a 3D convolutional autoencoder trained on normal MS data to localize anomalies without ground truth segmentation masks.
Training the SSL model to reconstruct these residual volumes enhances feature discrimination between normal and anomalous MS in the downstream classification task.
The method effectively utilizes available labelled healthy MS data reserved for downstream tasks, improving data efficiency.
Post-processing strategies and loss functions were investigated to optimize feature transferability for classification.
The approach differs from prior SSL methods by explicitly focusing on anomaly localization as part of the self-supervision task.
Clinical Implications
This SSL framework enables improved automated detection of maxillary sinus anomalies from 3D MRI without requiring extensive labeled datasets, potentially facilitating earlier and more accurate diagnosis. The method’s ability to leverage unlabelled data and localize anomalies may reduce reliance on expert annotation and improve clinical workflow efficiency in neuroradiological assessments.
Conclusion
The study demonstrates that self-supervised learning using anomaly localization via reconstruction residuals can significantly enhance classification of maxillary sinus anomalies from MRI. This approach offers a promising direction for leveraging unlabelled medical imaging data in clinical diagnostics.
References
Hamburg City Health Study (HCHS) 2020 -- Cranial MRI dataset
by Debayan Bhattacharya, Finn Behrendt, Benjamin Tobias Becker, Lennart Maack, Dirk Beyersdorff, Elina Petersen, Marvin Petersen, Bastian Cheng, Dennis Eggert, Christian Betz, Anna Sophie Hoffmann, Alexander Schlaefer
