Reducing the risk of hallucinations with interpretable deep learning models for low-dose CT denoising: comparative...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Radiomics

Background:

• Reducing CT radiation dose is crucial for patient safety, but it increases image noise and degrades clinical quality.
• Deep learning (DL) methods are increasingly used for low-dose CT (LDCT) denoising, yet concerns about image artifacts and interpretability persist.

Purpose of the Study:

• To assess the perceived quality, signal preservation, and radiomic feature preservation of LDCT volumes denoised by DL methods.
• To compare the clinical denoising performance of interpretable DL methods against classical deep neural networks.

Main Methods:

• Qualitative reader studies evaluated perceived image quality based on four criteria.
• Lesion detection/segmentation studies assessed the impact of denoising on signal detectability.
• Radiomic analysis compared quantitative and statistical similarity between denoised and standard-dose CT (SDCT) images.

Main Results:

• DL-denoised volumes were qualitatively inferior to SDCT volumes (p<0.05).
• Denoising did not significantly reduce lesion segmentation accuracy (p>0.05).
• Most denoised volumes yielded radiomics features statistically similar to SDCT volumes (p>0.05).

Conclusions:

• DL-denoised LDCT volumes exhibit lower perceived quality than SDCT but maintain comparable lesion detectability and radiomic feature integrity.
• These findings support the use of DL denoising for enhancing patient safety without compromising essential diagnostic information.