Self-supervised learning for classifying paranasal anomalies in the maxillary sinus.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Radiology
• Computational Pathology

Background:

• Paranasal anomalies are common findings in radiological screenings, presenting diverse morphologies.
• Supervised learning for anomaly classification requires extensive labeled datasets, which are often unavailable.
• Self-supervised learning (SSL) offers a promising alternative for learning from unlabeled data.

Purpose of the Study:

• To develop an SSL method tailored for classifying paranasal anomalies in the maxillary sinus (MS).
• To address the challenge of limited labeled data in training accurate anomaly detection models.
• To improve the efficiency and effectiveness of identifying diverse paranasal anomaly morphologies.

Main Methods:

• Utilized a 3D convolutional autoencoder (CAE) within an unsupervised anomaly detection (UAD) framework.
• Trained the CAE to reconstruct normal MS images, generating residual images highlighting anomalies.
• Employed a 3D convolutional neural network (CNN) for SSL task on residual images, followed by fine-tuning on labeled data.

Main Results:

• The proposed SSL technique demonstrated superior performance over generic SSL methods, particularly with limited annotated data.
• Achieved an area under the precision-recall curve (AUPRC) of 0.79 on 10% of the annotated dataset for MS anomaly classification.
• Outperformed BYOL (0.75), SimSiam (0.74), SimCLR (0.73), and SparK-based masked autoencoding (0.75) in AUPRC.

Conclusions:

• An SSL approach focusing on localizing paranasal anomalies is advantageous for differentiating normal from anomalous maxillary sinuses.
• The method shows significant potential for improving diagnostic accuracy in scenarios with scarce labeled medical imaging data.
• Code is available at https://github.com/mtec-tuhh/self-supervised-paranasal-anomaly for further research and application.