Addressing signal alterations induced in CT images by deep learning processing: A preliminary phantom study.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology

Background:

• Deep learning (DL) models are increasingly used for image processing in Computed Tomography (CT).
• Evaluating the performance and potential side effects of DL in CT is crucial for clinical application.

Purpose of the Study:

• To extensively evaluate the performance and potential side effects of DL processing on CT images.
• To compare two Convolutional Neural Networks (CNNs) for denoising and segmentation tasks.

Main Methods:

• Utilized autoencoder-based CNNs (encoder-decoder and UNet) for denoising and segmentation.
• Employed a phantom with iodinated contrast media for controlled image acquisition.
• Assessed CNN behavior using signal detection theory, radiological, conventional image quality, and radiomic features.

Main Results:

• The UNet model demonstrated superior performance in conventional quality parameters and spatial resolution compared to the encoder-decoder model.
• Radiomic analysis identified image features sensitive and insensitive to denoising processing by CNNs.
• CNNs were shown to modify image noise texture.

Conclusions:

• The evaluation approach effectively quantified differences in CNN behavior and image alterations.
• Complex DL networks achieving good performance may still undesirably alter image texture features.
• Careful quality assessment is necessary to understand the impact of DL on medical images.