LoMAE: Simple Streamlined Low-Level Masked Autoencoders for Robust, Generalized, and Interpretable Low-Dose CT...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Low-dose computed tomography (LDCT) reduces radiation exposure but degrades image quality with noise and artifacts.
• Transformer models show promise for LDCT image enhancement but require large paired datasets.
• Masked autoencoders (MAE) excel at feature extraction but are not directly applicable to low-level vision tasks like denoising.

Purpose of the Study:

• To develop a novel method for enhancing low-dose computed tomography (LDCT) image quality.
• To address the challenge of limited ground-truth data in clinical LDCT denoising.
• To improve the interpretability, robustness, and generalizability of transformer-based LDCT denoising models.

Main Methods:

• Redesigned the classical encoder-decoder learning model to create a streamlined low-level vision MAE (LoMAE).
• Applied LoMAE specifically for the LDCT denoising problem.
• Introduced MAE-GradCAM for visualizing and understanding the learning mechanisms of MAE/LoMAE.

Main Results:

• LoMAE significantly enhances the denoising performance of transformer models.
• The method substantially reduces the dependency on high-quality, ground-truth data for training.
• LoMAE demonstrated remarkable robustness and generalizability across various noise levels.

Conclusions:

• LoMAE offers a promising solution for improving LDCT image quality.
• The proposed method addresses key challenges in LDCT denoising, including interpretability and data dependency.
• LoMAE enhances model robustness and generalizability for clinical applications.