Applying Self-Supervised Learning to Image Quality Assessment in Chest CT Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Low-dose CT examinations require advanced reconstruction techniques with nonlinear properties.
• Assessing image quality necessitates task-based measures.
• The Hotelling observer (HO) offers optimal linear observer performance but is computationally complex.

Purpose of the Study:

• To develop a self-supervised learning (SSL)-based model observer for estimating HO performance in low-dose chest CT.
• To address the computational limitations of traditional HO methods.

Main Methods:

• A two-stage model combining a convolutional denoising auto-encoder (CDAE) for feature extraction and dimensionality reduction.
• A support vector machine (SVM) for classification.
• Signal detection tasks using simulated low-dose chest CT images with varying noise structures.

Main Results:

• The CDAE-based model achieved detection performance comparable to the HO.
• The proposed SSL approach outperformed single-layer neural networks (SLNN) and convolutional neural networks (CNN) with limited training data.
• The model demonstrated robustness across different noise structures.

Conclusions:

• The SSL-based model observer provides an effective and computationally efficient alternative to HO for low-dose CT image quality assessment.
• This method holds potential for optimizing low-dose CT protocols across various scanner platforms.
• The findings support the use of AI in enhancing medical imaging analysis.