Multiple instance ensembling for paranasal anomaly classification in the maxillary sinus.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Paranasal anomalies present diverse morphologies, challenging accurate classification by current AI models.
• Existing methods for paranasal anomaly classification are limited to single-anomaly identification.
• Accurate localization of maxillary sinuses (MS) in MRI scans is crucial for anomaly detection.

Purpose of the Study:

• To investigate the feasibility of using 3D convolutional neural networks (CNNs) for classifying healthy maxillary sinuses (MS) and those with polyps or cysts.
• To develop and evaluate a novel sampling technique for improving MS volume localization and dataset size.
• To enhance classification performance using a Multiple Instance Ensembling (MIE) prediction method.

Main Methods:

• Utilized 3D CNN architectures (ResNet, DenseNet) for MS anomaly classification.
• Implemented a novel sampling strategy to improve MS localization and augment training data.
• Applied Multiple Instance Ensembling (MIE) for enhanced prediction accuracy.

Main Results:

• Observed consistent improvements in classification performance across 3D CNN architectures.
• Sampling technique led to an average Area Under the Precision-Recall Curve (AUPRC) increase of 21.86 ± 11.92%.
• Combined sampling and MIE resulted in an average AUPRC increase of 28.86 ± 12.80% for ResNet and 9.85 ± 4.02% for DenseNet.

Conclusions:

• Sampling and MIE techniques effectively improve the generalizability of CNNs for paranasal anomaly classification.
• Demonstrated the feasibility of classifying anomalies within the maxillary sinuses using 3D CNNs.
• Proposed data augmentation via sampling and MIE as beneficial strategies for paranasal anomaly detection.