To design a self-supervised learning (SSL) task that improves the classification of paranasal anomalies in the maxillary sinus using unlabelled data, addressing challenges in clinical settings.
Key Findings:
The self-supervised method improved downstream classification of normal versus anomalous maxillary sinus, showcasing its effectiveness.
Anomaly localization was a primary focus of the self-supervision task, distinguishing it from previous methods that did not prioritize this aspect.
Varying the CAE training set significantly impacted downstream classification performance, highlighting the importance of dataset selection.
Interpretation:
The study demonstrates that self-supervised learning can effectively utilize unlabelled data to enhance the classification of paranasal anomalies, addressing the challenge of limited labelled datasets in clinical settings and paving the way for future research.
Limitations:
The study relies on the quality and representativeness of the unlabelled dataset, which may affect model performance.
Potential biases in the labelled dataset may affect the generalizability of the findings.
Conclusion:
The proposed self-supervised learning approach shows promise in improving the classification of paranasal anomalies, potentially leading to better patient outcomes through more accurate diagnoses.
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
