A deep-learning noise reduction algorithm outperforms the spatial filters previously required for bone SPECT on a high-speed whole-body 360° CZT-camera Authors: Achraf Bahloul, Franklin Rajadhas, Matthieu Doyen, Yechiel Lamash, Nathaniel Roth, Véronique Roch, Pierre-Yves Marie, Laetitia Imbert Published: 2025 ABSTRACT BackgroundSpatial filters are required to suppress the statistical noise of SPECT images but with an unavoidable smoothing effect that further decreases the SUV and contrast. This study assesses a deep-learning noise reduction (DLNR) algorithm, previously developed to further reduce bone SPECT recording time on a high-speed whole-body 360° CZT-camera, when used instead of, rather than in addition to, the conventional spatial filters (CSF) recommended for this camera.ResultsThe SUVmax of bone lesions (114 definite arthritis or metastasis lesions) and the resolution recovery coefficients of small and medium phantom spheres, were higher for DLNR than CSF or the combination of CSF plus DLNR (CSF-DLNR) (all p < 0.001), whereas the relative noises were lower for DLNR or CSF-DLNR, as compared with CSF (p < 0.001). Consequently, contrast-to-noise ratio (CNR) was dramatically higher for DLNR, as compared with CSF, and also CSF-DLNR, especially for small- and medium-sized structures. Compared with CSF, DLNR provided an almost two-fold CNR increase for the sphere and lesions in the range of one cm3. This dramatic CNR improvement was still documented when DLNR was compared with the median, kernel, Butterworth, or Gaussian filters used alone and set to provide an equivalent image noise reduction to DLNR on the phantom.ConclusionWhen used alone, this DLNR algorithm enhances the contrast-to-noise ratio and quantification of bone lesions, especially those of small or medium sizes. It outperforms conventional spatial filters and provides remarkable image quality for routine analysis of bone SPECT from the high-speed whole-body 360° CZT camera. However, further research and validation studies are still necessary before a widespread adoption in clinical practice.Key pointsQuestion: How does a deep-learning noise reduction algorithm, previously developed to further reduce bone SPECT recording times on a high-speed whole-body 360° CZT-camera, work when used instead of, rather than in addition to, the conventional spatial filters recommended for this camera.Pertinent findings: When used alone, this deep-learning noise reduction algorithm provides a high level of image denoising and better preserves the activities of small- to medium-sized bone structures and lesions than conventional spatial filters do, leading to a dramatic increase in the corresponding contrast-to-noise ratios.Implications for patient care: Such a deep-learning noise reduction algorithm could be used not only to reduce SPECT recording time when added to conventional spatial filters, but also to improve image quality and resolution when used alone.Trial registrationclinicaltrials.gov, NCT06782438, Registered 27 February 2025,https://clinicaltrials.gov/search?id=NCT06782438. Introduction Bone SPECT studies have reported the possibility for reduction in recording time and/or injected doses by applying deep-learning noise reduction (DLNR) algorithms [1–7]. We recently showed that whole-body bone SPECT recording times could be reduced to only 6 min using a newly developed DLNR algorithm for bone SPECT on a high-sensitivity whole-body 360° CZT-camera [8]. This was achieved whilst maintaining a high level of lesion detectability, albeit with somewhat decreased lesion quantification, as evidenced by slight declines in standardized uptake values (SUV) [8]. SPECT and PET quantifications of small and medium lesions can be underestimated due to partial volume effects. Such underestimation is further increased by the spatial filters required to suppress the statistical noise of SPECT images. Various spatial filters may be recommended for this purpose but with an unavoidable smoothing effect that further decreases spatial resolution and, consequently, decreases the SUV and contrast, especially for small lesions [9]. At the present time, it is not known to what extent conventional spatial filters (CSF) are still required when a DLNR algorithm is applied to bone SPECT. The present study assesses a DLNR algorithm, previously developed to further reduce bone SPECT recording time on a high-speed whole-body 360° CZT-camera, when used instead of, rather than in addition to, the conventional spatial filters (CSF) recommended for this camera. Materials and methods We retrospectively selected twenty patients who were referred to our Nuclear Medicine Department between November 2021 and May 2024 and underwent a dedicated whole-body bone SPECT imaging protocol on a high-sensitivity 360° cadmium zinc telluride (CZT) camera (VERITON-CT ® 400 series) for known or suspected polyarthritis (10 patients) or bone metastases (10 patients), each of whom had several bone lesions identified by experienced observers through a careful examination of conventional SPECT and CT images. This study was performed in accordance with the principles of the Declaration of Helsinki. The Ethics Committee of the CHRU Nancy approved the study protocol, which was released on the clinicaltrials.gov site under the identifier NCT06782438. The SPECT imaging protocol was also applied to a NEMA IEC body phantom (Data Spectrum Corporation, Durham, NC, USA). As previously described [10], we filled this phantom with a 99mTc solution providing a 10.7 spherical inserts/background activity ratio and associated a 300 HU density to the sphere contents. SPECT recording and reconstruction Delayed whole-body bone SPECT/CT acquisitions for patients started 3 to 4 h after the injection of 542 ± 28 MBq of [99mTc]Tc-hydroxy-diphosphonate (HDP), with CT recorded first [8]. The whole-body SPECT images were obtained according to the manufacturer’s recommendations [8] with (i) a total SPECT recording time ≤ 18 min, (ii) 6 to 7-bed positions recorded with focused mode and 140.0 ± 10.5 keV energy window, (iii) an ordered subset expectation maximization iterative reconstruction method (2 iterations, 8 subsets) correcting for the collimator response, with 2.46 mm cubic voxel size, and further corrections for scatter, attenuation, and partial volume, and (iv) conventional spatial filters (CSF) for decreasing the statistical noise -i.e., a 3 × 3 × 3 kernel inter-iteration filter with 0.2 ponderation factor and a 3 × 3 × 3 median post-reconstruction filter. The same, although single-bed only, SPECT imaging protocol was applied for the phantom, with the 2 min recording time subsequently shortened in the list mode data to achieve a total number of recorded counts in the range of those observed for the thoracoabdominal bed positions of our routine bone SPECT recordings (2 Mcounts). All patient and phantom SPECT images were additionally reconstructed with a DLNR algorithm applied instead of, or after the CSF (CSF-DLNR). As further detailed in Annex 1, this DLNR convolutional neural network is based on a 3D UNet architecture with residual units. The network was specifically trained on high-dose and low-dose image patches obtained with the VERITON-CT camera in another study group. The DLNR algorithm is part of a commercial software product dedicated to the VERITON-CT system, and it cannot be made publicly available. In an additional phantom experiment, the DLNR filter was compared with several of the current most commonly used SPECT filters, which were applied alone and adjusted to approach the DLNR noise reduction level - i.e., 3 × 3 × 3 Gaussian post-filter with a full width at half-maximum of 8 mm, Butterworth post-filter with a cut-off frequency of 0.25 cycles/pixel, and an order of 2, 5 × 5 × 5 median post-filter, and 3 × 3 × 3 inter-iteration kernel filter with 0.8 ponderation factor. SPECT image analysis As previously described [8], relative noise was measured with the standard-deviation (SD)/mean SUV ratio extracted from (i) two ellipsoid VOIs placed on the upper thighs for each patient and (ii) 12 VOIs of 37 mm each placed in the phantom background. Activities were quantified using: (i) SUVmax values for the metastatic or arthritis bone lesions and (ii) the maximal activity concentration recovery coefficient (RC) of the small (1.1 cm3), medium (5.6 cm3), and large (26.5 cm3) phantom spheres with the formula:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{\mathrm{Measured\:maximal\:activity}}{\mathrm{True\:activity}}$$\end{document} Contrast-to-noise (CNR) ratios were computed with the maximal activity of bone lesion or sphere (Max Activity) and the corresponding SD and mean background (BKG) levels [11, 12], as follows:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{\mathrm{Max}\text{}\mathrm{Activity\:-\:Mean\:BKG\:}}{\mathrm{SD\:BKG\:}}$$\end{document} The patient’s bone lesions were additionally regrouped into small, medium, and large categories. For this purpose, all patient lesions were displayed in ascending size order, and the category limits were adjusted so that the median size of each category was as close as possible to the size of the small, medium, and large phantom spheres (1.1 cm3, 5.6 cm3, and 26.5 cm3, respectively). Bone lesion sizes were determined with 38% SUVmax isocontours applied on the SPECT images reconstructed with the DLNR and no CSF. In a preliminary analysis, this method best predicted the phantom sphere volumes. Statistical analysis Quantitative data were presented as median [interquartile range], and paired comparisons were performed using Wilcoxon rank sum tests. A p-value < 0.05 was considered indicative of a significant difference and further adjusted for multiple comparisons by the Bonferroni method (i.e., with a statistical significance defined as p < 0.017 (0.05/3) for the pairwise comparisons shown in Table 1).Table 1Pairwise comparisons between images obtained with the recommended conventional spatial filters (CSF) alone, deep-learning noise reduction algorithm (DLNR) plus CSF, and DLNR alone, for the median values [interquartile range] of the relative noise extracted from the 20 patients’ whole-body SPECT and the SUVmax and contrast-to-noise ratio (CNR) of the 114 bone lesions classified into small, medium, and large size groups. Except for the relative noise comparison between CSF and CSF plus DLNR, all pairwise comparisons are significant -i.e., with Wilcoxon test p-values less than 0.017, which is the threshold for correcting multiple comparisons using the Bonferroni method CSF CSF + DLNR DLNR P valuesDLNR vs. CSF+DLNRCSF vs. DLNRCSF+DLNR vs. CSFRelative noise (%, n=20)29 [25–33]20 [17–22]21 [19–22]0.1800.001<0.001All lesions (n=114)SUVmax9.9 [7.1–13.8.1.8]9.2 [6.3–13.0.3.0]11.0 [8.1–15.4.1.4]<0.001<0.001<0.001CNR262 [158–376]331 [205–528]415 [256–634]<0.001<0.001<0.001Small lesions (n=12)SUVmax5.7 [4.5–11.4.5.4]5.4 [3–10.3.3]7.6 [4.2–14.6.2.6]0.0020.0080.002CNR163 [109–231]247 [146–310]337 [182–443]0.0100.0020.002Medium lesions(n=78)SUVmax9.9 [7.8–12.1.8.1]9.1 [6.8–11.2.8.2]11 [8.4–14.7.4.7]<0.001<0.001<0.001CNR266 [159–381]331 [223–511]415 [262–626]<0.001<0.001<0.001Large lesions (n=24)SUVmax12.4 [7.5–18.1.5.1]11.1 [6.7–16.8.7.8]14.3 [7.9–19.3.9.3]<0.0010.001<0.001CNR300 [179–462]456 [260–665]487 [326–767]<0.001<0.001<0.001 Table 1Pairwise comparisons between images obtained with the recommended conventional spatial filters (CSF) alone, deep-learning noise reduction algorithm (DLNR) plus CSF, and DLNR alone, for the median values [interquartile range] of the relative noise extracted from the 20 patients’ whole-body SPECT and the SUVmax and contrast-to-noise ratio (CNR) of the 114 bone lesions classified into small, medium, and large size groups. Except for the relative noise comparison between CSF and CSF plus DLNR, all pairwise comparisons are significant -i.e., with Wilcoxon test p-values less than 0.017, which is the threshold for correcting multiple comparisons using the Bonferroni method CSF CSF + DLNR DLNR P valuesDLNR vs. CSF+DLNRCSF vs. DLNRCSF+DLNR vs. CSFRelative noise (%, n=20)29 [25–33]20 [17–22]21 [19–22]0.1800.001<0.001All lesions (n=114)SUVmax9.9 [7.1–13.8.1.8]9.2 [6.3–13.0.3.0]11.0 [8.1–15.4.1.4]<0.001<0.001<0.001CNR262 [158–376]331 [205–528]415 [256–634]<0.001<0.001<0.001Small lesions (n=12)SUVmax5.7 [4.5–11.4.5.4]5.4 [3–10.3.3]7.6 [4.2–14.6.2.6]0.0020.0080.002CNR163 [109–231]247 [146–310]337 [182–443]0.0100.0020.002Medium lesions(n=78)SUVmax9.9 [7.8–12.1.8.1]9.1 [6.8–11.2.8.2]11 [8.4–14.7.4.7]<0.001<0.001<0.001CNR266 [159–381]331 [223–511]415 [262–626]<0.001<0.001<0.001Large lesions (n=24)SUVmax12.4 [7.5–18.1.5.1]11.1 [6.7–16.8.7.8]14.3 [7.9–19.3.9.3]<0.0010.001<0.001CNR300 [179–462]456 [260–665]487 [326–767]<0.001<0.001<0.001 Pairwise comparisons between images obtained with the recommended conventional spatial filters (CSF) alone, deep-learning noise reduction algorithm (DLNR) plus CSF, and DLNR alone, for the median values [interquartile range] of the relative noise extracted from the 20 patients’ whole-body SPECT and the SUVmax and contrast-to-noise ratio (CNR) of the 114 bone lesions classified into small, medium, and large size groups. Except for the relative noise comparison between CSF and CSF plus DLNR, all pairwise comparisons are significant -i.e., with Wilcoxon test p-values less than 0.017, which is the threshold for correcting multiple comparisons using the Bonferroni method Results Clinical study We analysed the whole-body SPECT recordings of 20 patients (9 women, median age: 72 [60–75] years), among whom 10 had bone metastases, and 10 had inflammatory arthritis lesions. A total of 114 bone lesion sites were considered (64 arthritis and 50 metastasis sites), and were grouped according to volume into three categories of small (n = 12, 1.2 [0.9–1.4] cm3), medium (n = 78, 4.8 [2.8–7.3] cm3) and large (n = 24, 24.3 [14.9–27.3] cm3) sized lesions. As detailed in Table 1, the SUVmax of all bone lesions were significantly higher for SPECT reconstructed with DLNR (11.0 [8.1–15.4]) than with CSF (9.9 [7.1–13.8], p < 0.001) or CSF-DLNR (9.2 [6.3–13.0], p < 0.001). Additionally, relative noise was comparable between DLNR (21 [19–22] %) and CSF-DLNR (20 [17–22] %, p = 0.18), but was clearly lower than with CSF (29 [25–33] %, p < 0.001 for CSF vs. DLNR or CSF-DLNR). Consequently, the CNR of bone lesions was markedly higher for DLNR (415 [256–634]) compared with CSF (262 [158–376], p = 0.001) and also with CSF-DLNR (331 [205–528], p < 0.001). As illustrated in Fig. 1 and detailed in Table 1, whatever the lesion size group, the median levels of SUVmax and CNR were significantly higher for DLNR than for CSF or CSF-DLNR (all p < 0.01). However, the differences were proportionally more critical for small-sized lesions than for large-sized ones. An almost two-fold CNR increase was documented between DLNR and CSF for the median of the small lesion group (i.e., the group with a 1.2 [0.9–1.4] cm3 volume) (Fig. 1). Fig. 1Comparisons between images obtained with the recommended conventional spatial filters (CSF) alone (red), deep-learning noise reduction algorithm (DLNR) plus CSF (purple), and DLNR alone (blue) for the median values of (A) the relative noise extracted from the 20 patients’ whole-body SPECT and (B), the SUVmax and contrast-to-noise ratio (CNR) of the 114 bone lesions classified into small (circles, n = 12, 1.2 [0.9–1.4] cm3), medium (triangles, n = 78, 4.8 [2.8–7.3] cm3) and large (squares, n = 24, 24.3 [14.9–27.3] cm3) size groups. In these three groups, the SUVmax and CNR associated with DLNR alone were significantly higher than those associated with CSF alone or CSF plus DLNR Comparisons between images obtained with the recommended conventional spatial filters (CSF) alone (red), deep-learning noise reduction algorithm (DLNR) plus CSF (purple), and DLNR alone (blue) for the median values of (A) the relative noise extracted from the 20 patients’ whole-body SPECT and (B), the SUVmax and contrast-to-noise ratio (CNR) of the 114 bone lesions classified into small (circles, n = 12, 1.2 [0.9–1.4] cm3), medium (triangles, n = 78, 4.8 [2.8–7.3] cm3) and large (squares, n = 24, 24.3 [14.9–27.3] cm3) size groups. In these three groups, the SUVmax and CNR associated with DLNR alone were significantly higher than those associated with CSF alone or CSF plus DLNR Representative patient SPECT images are displayed in Figs. 2, 3 and 4. Fig. 2T6 vertebral bone metastasis (arrow) observed in a 64-year-old woman with breast cancer on DLNR whole-body MIP (A), on enlarged thoracic MIPs (B), and coronal slices (C) with comparisons between images obtained with conventional spatial filters (CSF) alone, deep-learning noise reduction algorithm (DLNR) plus CSF, and DLNR alone. Note that with DLNR alone, all bone structures look sharper with higher contrast, and the T6 lesion exhibits higher SUVmax (blue numbers) T6 vertebral bone metastasis (arrow) observed in a 64-year-old woman with breast cancer on DLNR whole-body MIP (A), on enlarged thoracic MIPs (B), and coronal slices (C) with comparisons between images obtained with conventional spatial filters (CSF) alone, deep-learning noise reduction algorithm (DLNR) plus CSF, and DLNR alone. Note that with DLNR alone, all bone structures look sharper with higher contrast, and the T6 lesion exhibits higher SUVmax (blue numbers) Fig. 3An enthesopathy of the right trochiter (arrow) associated with a right acromioclavicular arthropathy in a 47-year-old man with ankylosing spondylitis on DLNR whole-body MIP (A), on enlarged right shoulder MIPs (B), and coronal slices (C) with comparisons between images obtained with conventional spatial filters (CSF) alone, deep-learning noise reduction algorithm (DLNR) plus CSF, and DLNR alone. Note that with DLNR alone, all bone structures look sharper with higher contrast, and the right trochiter lesion exhibits higher SUVmax (blue numbers) An enthesopathy of the right trochiter (arrow) associated with a right acromioclavicular arthropathy in a 47-year-old man with ankylosing spondylitis on DLNR whole-body MIP (A), on enlarged right shoulder MIPs (B), and coronal slices (C) with comparisons between images obtained with conventional spatial filters (CSF) alone, deep-learning noise reduction algorithm (DLNR) plus CSF, and DLNR alone. Note that with DLNR alone, all bone structures look sharper with higher contrast, and the right trochiter lesion exhibits higher SUVmax (blue numbers) Fig. 4MIP images obtained from hand volumes extracted from a whole-body high-speed SPECT recording (2 min for the bed position encompassing the hands) (A), as compared with a 5-min planar recording with the hands maintained in contact with the head of an Anger camera (Symbia, Siemens) (B), in a 68-year-old woman with arthrosis on the spine (L4-L5 discarthrosis), left shoulder and hands. Note the superior contrast of hand lesions (and the higher SUVmax of an interphalangeal joint in blue text) with DLNR alone than with CSF, DLNR plus CSF and the planar image MIP images obtained from hand volumes extracted from a whole-body high-speed SPECT recording (2 min for the bed position encompassing the hands) (A), as compared with a 5-min planar recording with the hands maintained in contact with the head of an Anger camera (Symbia, Siemens) (B), in a 68-year-old woman with arthrosis on the spine (L4-L5 discarthrosis), left shoulder and hands. Note the superior contrast of hand lesions (and the higher SUVmax of an interphalangeal joint in blue text) with DLNR alone than with CSF, DLNR plus CSF and the planar image Phantom study The phantom results obtained with the same SPECT reconstruction and filtering parameters as those used in the clinical study are detailed in Fig. 5. These results were very close to those observed in patients, with higher recovery coefficients and CNR for DLNR than for CSF or CSF-DLNR, especially for the smallest sphere of 1.1 cm3 (Fig. 5). An almost two-fold CNR increase was documented for DLNR alone vs. CSF alone for this smallest sphere (Fig. 5). The sole exception was the CNR of the large sphere, which was slightly higher for DLNR plus CSF than for DLNR alone. Fig. 5Same legend as that given in Fig. 1 but for the relative noise extracted from the phantom (A) and the recovery coefficient and contrast-to-noise ratio (CNR) of the small (circles, 1.1 cm3), medium, (triangles, 5.6 cm3) and large (squares, 26.5 cm3) phantom spheres (B). Axial sections of the SPECT phantom are also displayed (C). Note that higher CNR and RC (B) and superior sphere detection (C) are achieved with DLNR alone, especially for the smallest spheres Same legend as that given in Fig. 1 but for the relative noise extracted from the phantom (A) and the recovery coefficient and contrast-to-noise ratio (CNR) of the small (circles, 1.1 cm3), medium, (triangles, 5.6 cm3) and large (squares, 26.5 cm3) phantom spheres (B). Axial sections of the SPECT phantom are also displayed (C). Note that higher CNR and RC (B) and superior sphere detection (C) are achieved with DLNR alone, especially for the smallest spheres The dramatic improvement of the CNR from small and medium spheres was still documented when DLNR was compared with the median, kernel, Butterworth, or Gaussian filters used alone and set to provide a relative noise reduction that was in the range of the DLNR one (see Fig. 6). Fig. 6Same legend as that in Fig. 5 but for the comparison of DLNR (blue) with several conventional filters applied alone and adjusted to approach the DLNR noise reduction level (A) - i.e., 3 × 3 × 3 Gaussian post-filter with a FWHM of 8 mm (pink), Butterworth 2/0.25 post-filter (purple), 5 × 5 × 5 median post-filter (green), and 3 × 3 × 3 inter-iteration kernel filter (0.8 ponderation factor) (yellow). Note that higher CNR and RC (B) and superior sphere detection (C) are achieved with DLNR, especially for the smallest spheres Same legend as that in Fig. 5 but for the comparison of DLNR (blue) with several conventional filters applied alone and adjusted to approach the DLNR noise reduction level (A) - i.e., 3 × 3 × 3 Gaussian post-filter with a FWHM of 8 mm (pink), Butterworth 2/0.25 post-filter (purple), 5 × 5 × 5 median post-filter (green), and 3 × 3 × 3 inter-iteration kernel filter (0.8 ponderation factor) (yellow). Note that higher CNR and RC (B) and superior sphere detection (C) are achieved with DLNR, especially for the smallest spheres Discussion DLNR algorithms were initially developed to reduce the injected activities and recording times of PET and SPECT imaging when added to the conventional reconstruction and filtering processes. However, our results show that such algorithms could advantageously replace the spatial filters that are currently required for SPECT imaging on a high-speed whole-body 360° CZT-camera. Therefore, it is likely that such DLNR algorithms could be used not only to reduce SPECT recording time or injected activity, but also to improve image quality when used alone. Until now, image filtering has been consistently recommended for SPECT imaging of bone [13, 14] and other organs. This is largely due to the lower count statistics and, consequently, higher statistical noise of native SPECT images [9]. Unfortunately, the use of conventional SPECT filters is associated with a smoothing effect that reduces lesion SUVmax, image sharpness, and, ultimately, lesion detectability [9]. Our DLNR algorithm was specifically trained to suppress the relative increase in statistical noise associated with simulated decreases in SPECT image counts (i.e., while reducing the list mode data). Therefore, contrary to conventional spatial filters, such DLNR algorithms will likely act specifically on the statistical noise due to the low count of SPECT images. The present study shows that the DLNR-related denoising effect is associated with a higher preservation of lesion activities than conventional spatial filters, thereby improving the contrast-to-noise ratio and consequently, the detectability of bone lesions. Our phantom experiments evidenced these properties, in which the DLNR algorithm provided higher activities and CNR for the small and medium spheres compared to conventional filters used in association (Fig. 5) or alone (Fig. 6). The DLNR algorithm’s superiority was even more marked for patients than for phantom SPECT images, affecting all bone lesions regardless of size, although it was proportionally more marked for small- and medium-sized structures (Fig. 1). Indeed, considering this algorithm was trained on patient bone SPECT images, it is likely better adapted to the denoising of patient rather than phantom SPECT images. The DLNR-related CNR improvement was observed not only for bone lesions but also for normal small bone structures, as evidenced by the rib, vertebrae and hand images from Figs. 2, 3 and 4. Note that the interphalangeal joints in Fig. 4 have a volume close to 1 cm3, for which the CNR of phantom spheres and patient lesions is almost twofold higher with DLNR than CSF. The main limitations are the single-centre nature of the present study and a small study population of patients suffering from only bone metastasis or arthritis. Further research and validation studies are needed on more diverse populations at a larger, multi-centric scale. In addition, the assessment of the DLNR-related changes in diagnosis accuracy, diagnostic confidence and patient management, will require further studies, specifically designed according to the suspected or known pathology and disease severity. Finally, the current DLNR algorithm may not perform as effectively on SPECT images where the noise patterns and activity distributions are different from those of the bone SPECT images within the present study. This consideration is best illustrated by a more pronounced superiority of DLNR on patient lesions (Fig. 1) when compared to phantom spheres (Fig. 5). For other scanners and imaging protocols, however, specific DLNR algorithms may be elaborated by using the same methodology (i.e., generating reprocessed low-count datasets and retraining the model using similar hyperparameters). Unlike spatial filters, these DLNR algorithms will learn to reduce the statistical noise that affects current SPECT imaging, and as they are specifically elaborated for a noise reduction objective, the DLNR algorithms should be only minimally affected by smoothing effects, contrary to conventional spatial filters and in agreement with the present study results. Conclusion Our results show that, when used alone, the presented DLNR algorithm enhances the contrast/noise ratio and quantification of bone lesions, especially for small- and medium-sized structures. It outperforms conventional spatial filters and provides remarkable image quality for the routine analysis of bone SPECT from a high-speed whole-body 360° CZT camera. However, further research and validation studies are still necessary before a widespread adoption in clinical practice.