Low-Dose Abdominal CT Using a Deep Learning-Based Denoising Algorithm: A Comparison with CT Reconstructed with...

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Low-dose computed tomography (LD CT) is crucial for reducing radiation exposure.
• Traditional reconstruction methods like filtered back projection (FBP) and advanced modeled iterative reconstruction (ADMIRE) have limitations in noise reduction at low doses.
• Deep learning-based denoising algorithms (DLAs) offer a potential solution for improving LD CT image quality.

Purpose of the Study:

• To compare the image quality of LD CT images reconstructed using a DLA with those reconstructed using FBP and ADMIRE.
• To evaluate the effectiveness of DLAs in reducing noise while preserving diagnostic information.
• To assess the impact of training data dose on DLA performance.

Main Methods:

• A DLA was trained using routine-dose abdominal CT images as ground truth.
• Simulated LD CT images at 13%, 25%, and 50% of routine dose were generated and reconstructed using FBP, ADMIRE, and the trained DLA.
• Image quality was assessed using phantom studies (noise power spectrum, modulation transfer function) and patient data (mean image noise, qualitative analysis).

Main Results:

• LD CT images processed with DLAs exhibited significantly lower noise levels than those from LD-FBP and LD-ADMIRE (p < 0.001).
• The modulation transfer function (MTF) of LD-DLAs was lower than LD-ADMIRE and LD-FBP, indicating reduced spatial resolution.
• Qualitative analysis showed that LD-DLAs trained with a 50% simulated radiation dose (DLA-3) achieved the best overall image quality, comparable to LD-ADMIRE, with no additional artifacts.

Conclusions:

• DLAs effectively reduce noise in LD CT images compared to FBP and ADMIRE.
• While DLAs improve noise reduction, they may compromise spatial resolution.
• The DLA trained with a 50% simulated radiation dose provided the optimal balance of noise reduction and image quality.