CT-free PET for Pediatric Patients with Metal Implants

Le résumé fourni par la source

Metal implants in pediatric sarcoma patients pose an inherent challenge for deep-learning (DL) based attenuation and scatter correction (ASC) in PET imaging. To address this issue, we developed a maximum a priori joint activity and attenuation (MAP-AA) PET reconstruction technique to integrate DL-generated $\mu$ map as the regularization. We first trained an nnU-net using 287 pediatric PET/CT examinations between time-of-flight non-ASC (TOF NASC) PET images and corresponding CT-based $\mu$ map ($\mu^{\mathrm{CT}}$). In MAP-AA, the activity image $\lambda$ and the attenuation map $\mu$ were alternatively updated, with the DL model generated $\mu$ map ($\mu^{\mathrm{DL}}$) regularizing the transmission reconstruction together with a smoothness prior. We tested the proposed method on two pediatric sarcoma patients with metal implants in legs and a patient with no metal implant. The $\mu$ maps generated by the DL model ($\mu^{\mathrm{DL}}$) and from the MAP-AA algorithm without and with the $\mu^{\mathrm{DL}}$-based prior ($\mu^{\mathrm{MAP}}$ and $\mu^{\mathrm{DL} \text {-MAP }}$), following with the PET images reconstructed using them for ASC, were evaluated with the $\mu^{\mathrm{CT}}$ and $\mu^{\mathrm{CT}}$-based ASC PET image as the reference. Normalized mean squared error (NMSE, in $10^{-5}$) was calculated on the PET images and standardized uptake values (SUVs) were measured in the most focused lesion(s). Metal implants in two sarcoma patients were better recovered by the MAP-AA in both $\mu^{\mathrm{MAP}}$ and $\mu^{\mathrm{DL} \text {-MAP }}$ than in $\mu^{\mathrm{DL}}$. Cross-talk artifacts were better mitigated in $\mu^{\mathrm{DL} \text { MAP }}$ than those in $\mu^{\mathrm{MAP}}$. The overall similarity of $\lambda^{\mathrm{DL}-\mathrm{MAP}}$ to $\lambda^{\mathrm{CT}}$ is comparable to that of $\lambda^{\mathrm{DL}}$, with NMSE as $1.21,2.97$, and 1.32 for $\lambda^{\mathrm{DL}}, \lambda^{\mathrm{MAP}}$, and $\lambda^{\mathrm{DL}-\mathrm{MAP}}$, respectively. Furthermore, $\lambda^{\text {DL-MAP }}$ exhibits lower errors in metal-affected areas, with NMSE as 1.49, 2.67, and 1.25; $\operatorname{SUV}_{\text {max }}\left(\operatorname{SUV}_{\text {mean }}\right)$ relative difference as $4.8 \%(1.8 \%),-4.3 \%(-7.1 \%)$, and $2.1 \%(-0.7 \%)$, for $\lambda^{\mathrm{DL}}$, $\lambda^{\text {MAP }}$, and $\lambda^{\text {DL-MAP }}$, respectively. We conclude that the proposed DL regularized MAP-AA reconstruction leads to the potential of using CT-free PET imaging for pediatric sarcoma patients with metal implants.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.