. Author manuscript; available in PMC: 2026 Feb 18.

Published in final edited form as: IEEE Trans Med Imaging. 2026 Mar;45(3):1115–1126. doi: 10.1109/TMI.2025.3618247

TABLE II.

Effectiveness of the LeqMod in PET image denoising, across multiple vendors and varying levels of noise. Metrics include NRMSE, PSNR, SSIM, lesion SUV_mean bias, SUV_max bias, and TLG bias.

Scanner	Quadra						uExplorer						mCT

count level	1%	2%	5%	10%	25%	50%	1%	2%	5%	10%	25%	50%	5%	10%	20%

NRMSE (↓)
low-count	0.848	0.562	0.351	0.253	0.159	0.096	1.863	1.064	0.621	0.441	0.266	0.160	1.151	0.784	0.521

w/o	0.278	0.237	0.211	0.193	0.130	0.096	0.568	0.464	0.404	0.296	0.220	0.162	0.519	0.438	0.350
w	0.267 ^†	0.223 ^†	0.196 ^†	0.174 ^†	0.125 ^†	0.088 ^†	0.541 ^†	0.436 ^†	0.387 ^†	0.286 ^†	0.209 ^†	0.147 ^†	0.510 ^†	0.427 ^†	0.342 ^†

PSNR (↑)
low-count	53.496	53.491	53.145	54.601	57.767	61.920	42.709	44.554	47.718	50.062	53.933	58.116	44.487	46.325	48.944

w/o	52.405	53.883	55.424	56.640	58.903	61.413	46.154	48.224	50.352	52.048	54.431	57.937	47.306	48.809	50.731
w	52.762 ^†	54.051 ^†	55.743 ^†	56.850 ^†	59.190 ^†	62.220 ^†	46.835 ^†	48.656 ^†	50.745 ^†	52.285 ^†	54.968 ^†	58.038 ^†	47.285	48.899	50.857 ^†

SSIM (↑)
low-count	92.289	94.543	97.934	99.152	99.774	99.919	89.309	94.320	97.497	98.655	99.479	99.808	92.676	95.641	97.759

w/o	99.445	99.413	99.479	99.631	99.808	99.896	97.458	97.702	99.050	99.251	99.590	99.726	97.952	98.415	98.894
w	99.450 ^†	99.639 ^†	99.429	99.769 ^†	99.867 ^†	99.931 ^†	98.016 ^†	98.750 ^†	99.121 ^†	99.440 ^†	99.674 ^†	99.822 ^†	98.091 ^†	98.588 ^†	99.003 ^†

SUV_mean bias (%) (o)
low-count	−2.882	−0.444	1.189	1.703	1.488	1.052	−4.754	−0.780	1.180	1.378	0.917	0.815	−3.953	−2.137	−0.802

w/o	−20.948	−15.127	−9.445	−6.694	−3.858	−3.209	−34.051	−25.021	−15.913	−11.023	−7.318	−5.933	−25.517	−20.095	−13.199
w	−17.128 ^†	−12.840 ^†	−8.491 ^†	−5.711 ^†	−3.131 ^†	−1.557 ^†	−28.726 ^†	−20.591 ^†	−12.737 ^†	−8.471 ^†	−4.192 ^†	−2.494 ^†	−23.469 ^†	−17.493 ^†	−11.657 ^†

SUV_max bias (%) (o)
low-count	57.743	30.397	16.241	10.933	6.205	3.271	88.944	38.594	17.863	10.998	5.399	2.929	24.004	13.675	6.638

w/o	−26.473	−20.177	−13.468	−9.816	−6.474	−6.318	−43.518	−32.297	−20.356	−14.586	−10.081	−8.942	−28.522	−23.243	−16.480
w	−19.274 ^†	−15.108 ^†	−10.181 ^†	−6.959 ^†	−3.809 ^†	−2.546 ^†	−31.132 ^†	−23.759 ^†	−15.350 ^†	−10.463 ^†	−4.848 ^†	−2.985 ^†	−24.676 ^†	−18.803 ^†	−13.296 ^†

TLG bias (%) (o)
low-count	−1.963	−0.742	0.174	0.545	0.596	0.469	−2.362	0.127	1.382	1.368	0.921	0.769	−1.154	−0.476	−0.360

w/o	−5.142	−3.772	−2.446	−2.076	−1.311	−1.055	−13.643	−8.975	−5.463	−3.591	−2.495	−2.081	−8.386	−6.179	−4.172
w	−3.983 ^†	−3.261 ^†	−2.394	−1.671 ^†	−1.281	−0.692 ^†	−10.058 ^†	−6.332 ^†	−3.487 ^†	−2.321 ^†	−1.251 ^†	−0.744 ^†	−7.415 ^†	−5.169 ^†	−3.557 ^†

^†

P<0.05 based on the Wilcoxon signed-rank test between groups without (w/o) and with (w) the LeqMod.

(↑), (↓) and (o) represent expecting higher, lower, and zero-approaching values, respectively.