Restoration of Full Data from Sparse Data in Low-Dose Chest Digital Tomosynthesis Using Deep Convolutional Neural...

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Chest digital tomosynthesis (CDT) offers limited diagnostic information compared to CT and involves higher radiation doses than chest radiography, hindering clinical adoption.
• Reducing radiation dose in CDT to chest radiography levels is crucial for increasing its clinical utility.
• Generating high-quality images from limited data is essential due to the inherent trade-off between image quality and radiation dose in medical imaging.

Purpose of the Study:

• To investigate a novel approach for acquiring low-dose chest digital tomosynthesis (CDT) images using learning-based algorithms.
• To develop and evaluate a deep learning model for restoring image quality from sparse sampling data in CDT.
• To assess the radiation dose and diagnostic potential of the proposed low-dose CDT method.

Main Methods:

• A deep learning model based on end-to-end image translation using U-net was developed for image restoration.
• The model was trained using simulation and experimental CDT data, specifically reconstructing images from sparse to full sampling data (11 input, 81 output images).
• Effective radiation doses were measured using Monte Carlo simulations to evaluate the proposed method's dose profile.

Main Results:

• The deep learning model successfully restored image quality degraded by sparse sampling data.
• Quantitative evaluation using the Structure Similarity Index Measure (SSIM) showed an approximate 20% increase in SSIM.
• The effective dose for the proposed sparse sampling method was approximately 0.11 mSv, comparable to chest radiography (0.1 mSv).

Conclusions:

• The proposed learning-based reconstruction strategy effectively generates high-quality CDT images from sparse projection data.
• This method significantly improves image quality and reduces radiation dose, making CDT more clinically applicable.
• The approach offers a fast and potentially clinically viable alternative to traditional model-based reconstruction methods for sparse-data tomosynthesis.