Self-Supervised Multi-Scale Cropping and Simple Masked Attentive Predicting for Lung CT-Scan Anomaly Detection

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Anomaly detection in medical images, particularly lung CT scans, is challenging due to the need to identify subtle, localized irregularities without prior knowledge of anomaly types.
• Existing methods often rely on training with only normal data, which can limit their ability to detect diverse anomalies.

Purpose of the Study:

• To propose a self-supervised framework for learning robust representations of lung CT-scan images.
• To develop an effective out-of-distribution detector for identifying local and subtle anomalies in lung CT scans.
• To improve the accuracy and capabilities of anomaly detection in medical imaging.

Main Methods:

• A novel self-supervised framework utilizing multi-scale cropping and masked attentive prediction for lung CT-scan image representation learning.
• Introduction of CropMixPaste, a self-supervised augmentation task to generate density shadow-like anomalies, enhancing the detection of local irregularities.
• Development of a simple masked attentive predicting block (SMAPB) for refining local features by predicting masked context information.

Main Results:

• The proposed self-supervised framework successfully learns powerful representations for lung CT-scan anomaly detection.
• The CropMixPaste augmentation and SMAPB significantly improve the model's ability to detect local and subtle irregularities.
• Experimental results on real lung CT-scan datasets demonstrate the superiority of the proposed method over state-of-the-art techniques.

Conclusions:

• The developed self-supervised framework offers an effective approach for anomaly detection in lung CT scans.
• The combination of CropMixPaste and SMAPB enhances the detection of challenging local anomalies.
• This method provides a promising advancement for automated analysis and diagnosis in medical imaging.