Core Concepts
A self-supervised learning approach that focuses on localizing paranasal anomalies can enhance the classification of normal versus anomalous maxillary sinuses, particularly in scenarios with limited annotated data.
Abstract
The authors present a self-supervised learning (SSL) method for classifying paranasal anomalies in the maxillary sinus (MS). The key highlights are:
The SSL task involves training a 3D Convolutional Autoencoder (CAE) to reconstruct normal MS images, then using the CAE to generate residual volumes that serve as coarse anomaly localization masks for an unlabelled dataset. A 3D Convolutional Neural Network (CNN) is then trained to reconstruct these residual volumes.
The encoder part of the 3D CNN, trained on the SSL task, is then fine-tuned on a labelled dataset of normal and anomalous MS images for the downstream classification task.
The proposed SSL method outperforms generic SSL approaches like BYOL, SimSiam, and SimCLR, especially when the labelled training set is limited (e.g., 10% of the full dataset). It achieves an Area Under the Precision-Recall Curve (AUPRC) of 0.79 when trained on just 10% of the annotated data, compared to 0.75 for BYOL, 0.74 for SimSiam, 0.73 for SimCLR, and 0.75 for Masked Autoencoding using SparK.
The authors also investigate the impact of the CAE training set size on the downstream classification performance, finding that a larger normal MS dataset for CAE training leads to better anomaly localization and improved classification results.
The self-supervised approach proves advantageous, particularly in low-data scenarios, by leveraging unlabelled data to learn discriminative features for distinguishing normal and anomalous maxillary sinuses.
Stats
The labelled dataset consists of 1067 patients, with 489 exhibiting no pathologies in their left and right MS, and 578 having at least one MS presenting polyp, cyst or mucosal thickening pathology.
The unlabelled dataset consists of 1559 patient MRIs.
Quotes
"A self-supervised learning approach that inherently focuses on localizing paranasal anomalies proves to be advantageous, particularly when the subsequent task involves differentiating normal from anomalous maxillary sinuses."
"Our self-supervised task leverages available normal MS data, essential for supervised downstream task training."