Evaluation of deep learning-based reconstruction models on non-TOF BGO PET/CT : impact of acquisition times and BSREM penalization factors on lesion detectability and SNR
Stenvall, Anna LU ; Minarik, David LU ; Trägårdh, Elin LU
and Kvernby, Sofia
LU
(2026)
In EJNMMI Physics
13(1).
- Abstract
Background: New long field-of-view (FOV) PET scanners using bismuth germanate (BGO) detectors without time-of-flight (TOF) capability are now available. These systems incorporate deep learning-based TOF (DLb-TOF) models to compensate for the absence of TOF. There is a lack of studies systematically investigating the optimal balance between signal to noise ratio and lesion detectability across a broader range of acquisition times and β-values for these DLb-TOF models. This study aims to evaluate the trade-off between acquisition time, signal-to-noise ratio (SNR) and lesion detectability to guide optimization of clinical protocol. Materials and methods: Twenty patients referred for a clinical [18F]fluorodeoxyglucose (FDG) PET... (More)
Background: New long field-of-view (FOV) PET scanners using bismuth germanate (BGO) detectors without time-of-flight (TOF) capability are now available. These systems incorporate deep learning-based TOF (DLb-TOF) models to compensate for the absence of TOF. There is a lack of studies systematically investigating the optimal balance between signal to noise ratio and lesion detectability across a broader range of acquisition times and β-values for these DLb-TOF models. This study aims to evaluate the trade-off between acquisition time, signal-to-noise ratio (SNR) and lesion detectability to guide optimization of clinical protocol. Materials and methods: Twenty patients referred for a clinical [18F]fluorodeoxyglucose (FDG) PET scan were included. Each patient received 3.5 MBq/kg of [18F]FDG and underwent a whole-body PET acquisition (120 s/bed) on a digital BGO PET/CT (32 cm FOV) 60 min post-injection. Data were reconstructed into images (384 × 384 matrix) representing different acquisition times (120 s, 90 s, 60 s, 45 s, 30 s and 15 s) using BSREM with β-values ranging from 50 to 1100. Three DLb-TOF models (Low, Medium, High) were applied. Volumes of interest were placed in the liver and two avid lesions per patient. SNR were calculated as SUVmeanliver/SDliver and detectability were calculated as SUVpeaktumor/SUVpeakliver. Results: SNR increased with longer acquisition times and higher β-values. DLb-TOF models improved SNR across all settings, with the Low DLb-TOF model producing the largest increase. Lesion detectability depended on the acquisition time and β-value. At longer acquisition times (120 s, 90 s), β100 provided the highest detectability, while shorter times (60–15 s) required higher β-value (β300) for optimal detectability. Among DLb-TOF models, the High model gave the best detectability overall, though the Low model performed better at lower β-values. Conclusion: SNR increased with higher β-values, longer acquisition times, and DLb-TOF application. Lesion detectability, defined as the ratio of SUVpeak in the lesion to SUVpeak in the liver, depended on the β-value, acquisition time, and the DLb-TOF model used. The Low DLb-TOF model had the best SNR but at the expense of detectability. The optimal parameters for the evaluated BGO PET/CT system, balancing SNR and lesion detectability within a clinical reasonable acquisition time, were 60–90 s with β-values of 500–300, in combination with the Medium DLb-TOF model, when 3.5 MBq/kg [18F]FDG was administered.
(Less)
- author
- Stenvall, Anna
LU
; Minarik, David
LU
; Trägårdh, Elin
LU
and Kvernby, Sofia
LU
- organization
- publishing date
- 2026-12
- type
- Contribution to journal
- publication status
- published
- subject
- keywords
- BSREM, Deep-learning, PET/CT, Time of flight
- in
- EJNMMI Physics
- volume
- 13
- issue
- 1
- article number
- 74
- publisher
- Springer Science and Business Media B.V.
- external identifiers
-
- pmid:42437412
- scopus:105044548283
- ISSN
- 2197-7364
- DOI
- 10.1186/s40658-026-00919-8
- language
- English
- LU publication?
- yes
- additional info
- Publisher Copyright: © The Author(s) 2026.
- id
- b63cd661-a676-4ceb-be5a-92f02b85c402
- date added to LUP
- 2026-10-02 15:15:26
- date last changed
- 2026-10-03 03:00:03
@article{b63cd661-a676-4ceb-be5a-92f02b85c402,
abstract = {{<p>Background: New long field-of-view (FOV) PET scanners using bismuth germanate (BGO) detectors without time-of-flight (TOF) capability are now available. These systems incorporate deep learning-based TOF (DLb-TOF) models to compensate for the absence of TOF. There is a lack of studies systematically investigating the optimal balance between signal to noise ratio and lesion detectability across a broader range of acquisition times and β-values for these DLb-TOF models. This study aims to evaluate the trade-off between acquisition time, signal-to-noise ratio (SNR) and lesion detectability to guide optimization of clinical protocol. Materials and methods: Twenty patients referred for a clinical [<sup>18</sup>F]fluorodeoxyglucose (FDG) PET scan were included. Each patient received 3.5 MBq/kg of [<sup>18</sup>F]FDG and underwent a whole-body PET acquisition (120 s/bed) on a digital BGO PET/CT (32 cm FOV) 60 min post-injection. Data were reconstructed into images (384 × 384 matrix) representing different acquisition times (120 s, 90 s, 60 s, 45 s, 30 s and 15 s) using BSREM with β-values ranging from 50 to 1100. Three DLb-TOF models (Low, Medium, High) were applied. Volumes of interest were placed in the liver and two avid lesions per patient. SNR were calculated as SUVmean<sub>liver</sub>/SD<sub>liver</sub> and detectability were calculated as SUVpeak<sub>tumor</sub>/SUVpeak<sub>liver</sub>. Results: SNR increased with longer acquisition times and higher β-values. DLb-TOF models improved SNR across all settings, with the Low DLb-TOF model producing the largest increase. Lesion detectability depended on the acquisition time and β-value. At longer acquisition times (120 s, 90 s), β100 provided the highest detectability, while shorter times (60–15 s) required higher β-value (β300) for optimal detectability. Among DLb-TOF models, the High model gave the best detectability overall, though the Low model performed better at lower β-values. Conclusion: SNR increased with higher β-values, longer acquisition times, and DLb-TOF application. Lesion detectability, defined as the ratio of SUV<sub>peak</sub> in the lesion to SUV<sub>peak</sub> in the liver, depended on the β-value, acquisition time, and the DLb-TOF model used. The Low DLb-TOF model had the best SNR but at the expense of detectability. The optimal parameters for the evaluated BGO PET/CT system, balancing SNR and lesion detectability within a clinical reasonable acquisition time, were 60–90 s with β-values of 500–300, in combination with the Medium DLb-TOF model, when 3.5 MBq/kg [<sup>18</sup>F]FDG was administered.</p>}},
author = {{Stenvall, Anna and Minarik, David and Trägårdh, Elin and Kvernby, Sofia}},
issn = {{2197-7364}},
keywords = {{BSREM; Deep-learning; PET/CT; Time of flight}},
language = {{eng}},
number = {{1}},
publisher = {{Springer Science and Business Media B.V.}},
series = {{EJNMMI Physics}},
title = {{Evaluation of deep learning-based reconstruction models on non-TOF BGO PET/CT : impact of acquisition times and BSREM penalization factors on lesion detectability and SNR}},
url = {{http://dx.doi.org/10.1186/s40658-026-00919-8}},
doi = {{10.1186/s40658-026-00919-8}},
volume = {{13}},
year = {{2026}},
}