Towards ultra-low-dose CT for detecting pulmonary nodules using DenseNet

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Ultra-low-dose CT (ULCT) is crucial for lung cancer screening but suffers from image noise, hindering nodule detection.
• Reducing radiation dose is vital due to the linear no-threshold model, which posits no safe radiation level.

Purpose of the Study:

• To investigate the feasibility of using DenseNet, a deep learning model, for noise suppression in ULCT images.
• To assess if DenseNet can improve image quality and nodule detectability in ULCT for lung cancer screening.

Main Methods:

• DenseNet was trained on CT images with varying radiation doses.
• The model was tested on 14 patients (7 with solid nodules, 7 with subsolid nodules) not included in training.
• Image quality was evaluated using Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), Contrast-to-Noise Ratio (CNR), and subjective scoring.

Main Results:

• Denoising with DenseNet significantly improved RMSE and PSNR values.
• Lung nodules were more easily distinguishable in denoised ULCT images, supported by improved CNR and subjective assessments.
• No statistically significant differences were found between full-dose CT and denoised ULCT in evaluating anatomical structures.

Conclusions:

• DenseNet is a viable approach for effectively reducing image noise in ULCT scans.
• This technique shows promise for enhancing lung cancer screening by improving nodule detection without compromising diagnostic accuracy.
• Further dose reduction in CT imaging remains a critical area of research in radiology.