CNN reduces metal artifacts in ultra-high-resolution photon-counting CT of ear implants

A convolutional neural network improved artifact index by 83.2% and reduced noise by 46.8% (p<0.001) on ultra-high-resolution photon-counting CT of stapes prostheses, cochlear implants, and a sinus stent, preserving metal dimensions.

• Mean-absolute-error loss with artifact amplification scalar 20 simultaneously preserved bone CT number and reduced artifacts, avoiding the trade-off seen with mean-squared-error training.
• Artifact reduction was greatest for stapes prostheses (91.1%) versus cochlear implants (81.3%) and a transverse sinus stent (72.9%).
• Metal implant dimensions measured by full-width at half-maximum were not significantly different between network input and output (p=0.67).

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.