Pediatric evaluations for deep learning CT denoising.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Computational Phantoms

Background:

• Deep learning (DL) CT denoising models enhance image quality at lower radiation doses.
• These models are typically trained on adult data, raising concerns for pediatric applications.
• Pediatric anatomy varies significantly, necessitating subgroup-specific performance evaluations.

Purpose of the Study:

• To develop and assess a framework for evaluating DL CT denoising performance in pediatric and adult-sized patients.
• To create computer-simulated image quality (IQ) phantoms representing diverse pediatric body sizes.

Main Methods:

• Simulated CT images using pediatric-sized (newborn to 18 years) and adult-sized IQ phantoms (CatPhan 600, MITA-LCD).
• Evaluated a DL denoiser (REDCNN trained on adult data) on simulated adult and pediatric images.
• Assessed image quality changes: noise, sharpness, CT number accuracy, and low contrast detectability.
• Validated findings using anthropomorphic pediatric XCAT phantoms.

Main Results:

• Adult-trained DL denoising model performance significantly decreased in smaller pediatric-sized phantoms.
• Noise reduction dropped by over 60% in smaller phantoms due to altered noise textures from smaller fields of view (FOV).
• Validation with XCAT phantoms confirmed noise reduction trends observed with IQ phantoms.

Conclusions:

• A novel framework using pediatric-sized IQ phantoms enables effective evaluation of DL denoising models for pediatric subgroups.
• Adult-trained DL denoisers exhibit poor generalization to smaller pediatric patient sizes.
• FOV differences between adult and pediatric protocols contribute to performance disparities, highlighting the need for tailored DL models or evaluation frameworks.