Margin-aware intraclass novelty identification for medical images

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Existing anomaly detection methods struggle with intraclass variations in medical images, limiting their ability to identify rare diseases.
• Medical image novelty detection is challenged by subtle differences within the same class and a lack of diverse training data.

Purpose of the Study:

• To develop a robust anomaly detection model capable of identifying novel or rare medical conditions within images, specifically addressing intraclass variations.
• To propose a hybrid approach combining AutoEncoder-based and classifier-based methods for enhanced novelty detection in medical imaging.

Main Methods:

• Introduced TEND (Transformation-based Embedding Learning for Novelty Detection), a two-stage hybrid model.
• Stage 1: Unsupervised learning of in-distribution embeddings using an AutoEncoder.
• Stage 2: Training a discriminative classifier with a margin-aware objective to differentiate normal and transformed data, generating an anomaly score.

Main Results:

• TEND demonstrated competitive performance against state-of-the-art methods on three public medical imaging datasets.
• Experiments included one-vs-rest and rest-vs-one setups, validating TEND's effectiveness on intraclass out-of-distribution data.
• Qualitative examples confirmed TEND's capability in identifying challenging medical image anomalies.

Conclusions:

• The proposed TEND model effectively identifies intraclass out-of-distribution medical images in an unsupervised manner.
• TEND can discover previously unseen medical image classes and assist in abnormal data screening for downstream tasks.
• The study provides open-source code for the TEND model to facilitate further research and application.