Nikolas D. Schnellbaecher
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Nikolas D. Schnellbaecher has a total of 36 co-authors. Here are the top co-authors by number of shared publications:
- Michael Grass
Innovative Technologies, Philips, Hamburg, Germany
3 shared publications - Anna Falk Delgado
Department of Clinical Neuroscience, Karolinska Institutet, Stockholm, Sweden
1 shared publication Christian Wuelker
Innovative Technologies, Philips, Hamburg, Germany
1 shared publication- Christopher Schuppert
Department of Diagnostic and Interventional Radiology, Medical Center - University of Freiburg, Faculty of Medicine, University of Freiburg, Freiburg, Germany
1 shared publication - Dominika Suchá
Department of Radiology and Nuclear Medicine, University Medical Center Utrecht, Utrecht, The Netherlands
1 shared publication - Erik Hummel
Philips (Netherlands), Best, Netherlands
1 shared publication
Latest Publications
- Open Access
- 24 August 2026
- Published by Springer Nature in Neuroradiology
- Vol. 68 (9) , 3083-3091
Brain cone-beam CT image quality improvement using a deep-learning-based denoising method: a multicenter retrospective study
Purpose: Deep learning (DL) denoising may improve cone-beam CT (CBCT) image quality for point-of-care stroke assessment in the interventional suite. The purpose of this study was to evaluate the impact of a DL-based denoising algorithm on objective and subjective image quality in brain CBCT using both standard circular and advanced dual-axis trajectories. Methods: We retrospectively analyzed 20 noncontrast brain CBCT acquisitions (Karolinska: 10 standard circular; St Michael's: 10 dual-axis). A DL-based denoising algorithm was applied at three strengths (Minimal, Medium, High) and compared to standard images with no additional denoising. Objective metrics (noise, signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), artifact indices) were measured using standardized ROIs. Six experts rated subjective image quality on 5-point Likert scales. Paired tests with Bonferroni correction were used for comparisons. Results: High-level DL denoising significantly improved all objective metrics (noise, SNR, CNR, artifact indices) for both thin and thick slices (all p <.001). It doubled gray–white matter CNR (thin slices: 2.31 vs. 1.08), reduced noise, and improved subcalvarial and posterior fossa artifact indices. Subjectively, high-level denoising yielded higher median ratings for noise, texture, sharpness, brain parenchyma visualization, CSF spaces, and confidence in assessing ischemia and hemorrhage (all p <.001). Improvements were consistent for both acquisition techniques, and perceived artifact severity did not differ ( p >.99). Inter-reader agreement was substantial. Conclusion: The DL-based denoising algorithm significantly improved objective and subjective brain CBCT image-quality; no difference in perceived artifact severity was detected. These findings support further evaluation of deep learning-enhanced CBCT denoising for brain imaging in the interventional suite.
- 24 August 2026
- Published by Springer Nature in Neuroradiology
- Vol. 68 (9) , 3083-3091
Open AccessBrain cone-beam CT image quality improvement using a deep-learning-based denoising method: a multicenter retrospective study
Purpose: Deep learning (DL) denoising may improve cone-beam CT (CBCT) image quality for point-of-care stroke assessment in the interventional suite. The purpose of this study was to evaluate the impact of a DL-based denoising algorithm on objective and subjective image quality in brain CBCT using both standard circular and advanced dual-axis trajectories. Methods: We retrospectively analyzed 20 noncontrast brain CBCT acquisitions (Karolinska: 10 standard circular; St Michael's: 10 dual-axis). A DL-based denoising algorithm was applied at three strengths (Minimal, Medium, High) and compared to standard images with no additional denoising. Objective metrics (noise, signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), artifact indices) were measured using standardized ROIs. Six experts rated subjective image quality on 5-point Likert scales. Paired tests with Bonferroni correction were used for comparisons. Results: High-level DL denoising significantly improved all objective metrics (noise, SNR, CNR, artifact indices) for both thin and thick slices (all p <.001). It doubled gray–white matter CNR (thin slices: 2.31 vs. 1.08), reduced noise, and improved subcalvarial and posterior fossa artifact indices. Subjectively, high-level denoising yielded higher median ratings for noise, texture, sharpness, brain parenchyma visualization, CSF spaces, and confidence in assessing ischemia and hemorrhage (all p <.001). Improvements were consistent for both acquisition techniques, and perceived artifact severity did not differ ( p >.99). Inter-reader agreement was substantial. Conclusion: The DL-based denoising algorithm significantly improved objective and subjective brain CBCT image-quality; no difference in perceived artifact severity was detected. These findings support further evaluation of deep learning-enhanced CBCT denoising for brain imaging in the interventional suite.
- 8 July 2026
- Published by Wolters Kluwer Health in Journal of Computer Assisted Tomography
- Vol. 50 (6)
Deep Learning-Based Deblurring for Computed Tomography and a Phantom Evaluation for Coronary Artery Calcium Detectability
Introduction: This study aimed to develop and evaluate a deep learning (DL) deblurring algorithm for computed tomography (CT). Phantom testing assessed whether DL deblurring could improve CT image quality and coronary artery calcium (CAC) detectability. Materials and Methods: The algorithm consisted of 2 sequential convolutional neural networks (CNNs): one for denoising and one for deblurring. After deblurring, the estimated noise from the first model was reintroduced to preserve image texture. The CNN denoising model was trained to map high-noise images to routine low-noise ones. The CNN deblurring model was trained using natural images—blurred inputs paired with sharp targets. To assess performance, an anthropomorphic phantom with 100 small calcifications was scanned on a CT system using a CAC scoring protocol. Data were acquired at clinical and high doses, then reconstructed with and without DL deblurring, using 25% of the maximum denoising strength. Detectability was defined as the ability to calculate an Agatston score (≥3 adjacent voxels >130 HU). Results: At high dose, CAC detectability increased from 39% (standard reconstruction) to 48% with deblurring. Similar detectability (39%) was achieved at the routine dose with deblurring, representing a 6% improvement compared with the original routine dose data. Discussion: Quantitative results from phantom testing demonstrate that DL deblurring can improve image quality, based on a clinically relevant, task-based CAC detection metric at clinical dose levels.
- 8 July 2026
- Published by Wolters Kluwer Health in Journal of Computer Assisted Tomography
- Vol. 50 (6)
Deep Learning-Based Deblurring for Computed Tomography and a Phantom Evaluation for Coronary Artery Calcium Detectability
Introduction: This study aimed to develop and evaluate a deep learning (DL) deblurring algorithm for computed tomography (CT). Phantom testing assessed whether DL deblurring could improve CT image quality and coronary artery calcium (CAC) detectability. Materials and Methods: The algorithm consisted of 2 sequential convolutional neural networks (CNNs): one for denoising and one for deblurring. After deblurring, the estimated noise from the first model was reintroduced to preserve image texture. The CNN denoising model was trained to map high-noise images to routine low-noise ones. The CNN deblurring model was trained using natural images—blurred inputs paired with sharp targets. To assess performance, an anthropomorphic phantom with 100 small calcifications was scanned on a CT system using a CAC scoring protocol. Data were acquired at clinical and high doses, then reconstructed with and without DL deblurring, using 25% of the maximum denoising strength. Detectability was defined as the ability to calculate an Agatston score (≥3 adjacent voxels >130 HU). Results: At high dose, CAC detectability increased from 39% (standard reconstruction) to 48% with deblurring. Similar detectability (39%) was achieved at the routine dose with deblurring, representing a 6% improvement compared with the original routine dose data. Discussion: Quantitative results from phantom testing demonstrate that DL deblurring can improve image quality, based on a clinically relevant, task-based CAC detection metric at clinical dose levels.
- Open Access
- 25 August 2025
- Published by MDPI AG in Tomography
- Vol. 11 (9) , 94
Performance of a Deep Learning Reconstruction Method on Clinical Chest–Abdomen–Pelvis Scans from a Dual-Layer Detector CT System
Objective: The objective of this study was to compare the performance and robustness of a deep learning reconstruction method against established alternatives for soft tissue CT image reconstruction. Materials and Methods: Images were generated from portal venous phase chest–abdomen–pelvis CT scans (n = 99) acquired on a dual-layer spectral detector CT using filtered back projection, iterative model reconstruction (IMR), and deep learning reconstruction (DLR) with three parameter settings, namely ‘standard’, ‘sharper’, and ‘smoother’. Experienced raters performed a quantitative assessment by considering attenuation stability and image noise levels in ten representative structures across all reconstruction methods, as well as a qualitative assessment using a four-point Likert scale (1 = poor, 2 = fair, 3 = good, 4 = excellent) for their overall perception of ‘smoother’ DLR and IMR images. One scan was excluded due to cachexia, which limited the quantitative measurements. Results: The inter-rater reliability for quantitative measurements ranged from moderate to excellent (r = 0.63–0.96). Attenuation values did not differ significantly between reconstruction methods except for DLR against IMR in the psoas muscle (mean + 3.0 HU, p < 0.001). Image noise levels differed significantly between reconstruction methods for all structures (all p < 0.001) and were lower than FBP with any DLR parameter setting. Image noise levels with ‘smoother’ DLR were predominantly lower than or equal to IMR, while they were higher with ‘standard’ DLR and ‘sharper’ DLR. The ‘smoother’ DLR images received a higher mean rating for overall image quality than the IMR images (3.7 vs. 2.3, p < 0.001). Conclusions: ‘Smoother’ DLR images were perceived by experienced readers as having improved quality compared to FBP and IMR while also exhibiting objectively lower or equivalent noise levels.
- 25 August 2025
- Published by MDPI AG in Tomography
- Vol. 11 (9) , 94
Open AccessPerformance of a Deep Learning Reconstruction Method on Clinical Chest–Abdomen–Pelvis Scans from a Dual-Layer Detector CT System
Objective: The objective of this study was to compare the performance and robustness of a deep learning reconstruction method against established alternatives for soft tissue CT image reconstruction. Materials and Methods: Images were generated from portal venous phase chest–abdomen–pelvis CT scans (n = 99) acquired on a dual-layer spectral detector CT using filtered back projection, iterative model reconstruction (IMR), and deep learning reconstruction (DLR) with three parameter settings, namely ‘standard’, ‘sharper’, and ‘smoother’. Experienced raters performed a quantitative assessment by considering attenuation stability and image noise levels in ten representative structures across all reconstruction methods, as well as a qualitative assessment using a four-point Likert scale (1 = poor, 2 = fair, 3 = good, 4 = excellent) for their overall perception of ‘smoother’ DLR and IMR images. One scan was excluded due to cachexia, which limited the quantitative measurements. Results: The inter-rater reliability for quantitative measurements ranged from moderate to excellent (r = 0.63–0.96). Attenuation values did not differ significantly between reconstruction methods except for DLR against IMR in the psoas muscle (mean + 3.0 HU, p < 0.001). Image noise levels differed significantly between reconstruction methods for all structures (all p < 0.001) and were lower than FBP with any DLR parameter setting. Image noise levels with ‘smoother’ DLR were predominantly lower than or equal to IMR, while they were higher with ‘standard’ DLR and ‘sharper’ DLR. The ‘smoother’ DLR images received a higher mean rating for overall image quality than the IMR images (3.7 vs. 2.3, p < 0.001). Conclusions: ‘Smoother’ DLR images were perceived by experienced readers as having improved quality compared to FBP and IMR while also exhibiting objectively lower or equivalent noise levels.
- 28 March 2022
- Published by SPIE-Intl Soc Optical Eng in Journal of Medical Imaging
- Vol. 9 (02) , 025001-20
Cascaded learning in intravascular ultrasound: coronary stent delineation in manual pullbacks
Purpose: Implanting stents to re-open stenotic lesions during percutaneous coronary interventions is considered a standard treatment for acute or chronic coronary syndrome. Intravascular ultrasound (IVUS) can be used to guide and assess the technical success of these interventions. Automatically segmenting stent struts in IVUS sequences improves workflow efficiency but is non-trivial due to a challenging image appearance entailing manifold ambiguities with other structures. Manual, ungated IVUS pullbacks constitute a challenge in this context. We propose a fully data-driven strategy to first longitudinally detect and subsequently segment stent struts in IVUS frames. Approach: A cascaded deep learning approach is presented. It first trains an encoder model to classify frames as “stent,” “no stent,” or “no use.” A segmentation model then delineates stent struts on a pixel level only in frames with a stent label. The first stage of the cascade acts as a gateway to reduce the risk for false positives in the second stage, the segmentation, which is trained on a smaller and difficult-to-annotate dataset. Training of the classification and segmentation model was based on 49,888 and 1826 frames of 74 sequences from 35 patients, respectively. Results: The longitudinal classification yielded Dice scores of 92.96%, 82.35%, and 94.03% for the classes stent, no stent, and no use, respectively. The segmentation achieved a Dice score of 65.1% on the stent ground truth (intra-observer performance: 75.5%) and 43.5% on all frames (including frames without stent, with guidewires, calcium, or without clinical use). The latter improved to 49.5% when gating the frames by the classification decision and further increased to 57.4% with a heuristic on the plausible stent strut area. Conclusions: A data-driven strategy for segmenting stents in ungated, manual pullbacks was presented—the most common and practical scenario in the time-critical clinical workflow. We demonstrated a mitigated risk for ambiguities and false positive predictions.
- 28 March 2022
- Published by SPIE-Intl Soc Optical Eng in Journal of Medical Imaging
- Vol. 9 (02) , 025001-20
Cascaded learning in intravascular ultrasound: coronary stent delineation in manual pullbacks
Purpose: Implanting stents to re-open stenotic lesions during percutaneous coronary interventions is considered a standard treatment for acute or chronic coronary syndrome. Intravascular ultrasound (IVUS) can be used to guide and assess the technical success of these interventions. Automatically segmenting stent struts in IVUS sequences improves workflow efficiency but is non-trivial due to a challenging image appearance entailing manifold ambiguities with other structures. Manual, ungated IVUS pullbacks constitute a challenge in this context. We propose a fully data-driven strategy to first longitudinally detect and subsequently segment stent struts in IVUS frames. Approach: A cascaded deep learning approach is presented. It first trains an encoder model to classify frames as “stent,” “no stent,” or “no use.” A segmentation model then delineates stent struts on a pixel level only in frames with a stent label. The first stage of the cascade acts as a gateway to reduce the risk for false positives in the second stage, the segmentation, which is trained on a smaller and difficult-to-annotate dataset. Training of the classification and segmentation model was based on 49,888 and 1826 frames of 74 sequences from 35 patients, respectively. Results: The longitudinal classification yielded Dice scores of 92.96%, 82.35%, and 94.03% for the classes stent, no stent, and no use, respectively. The segmentation achieved a Dice score of 65.1% on the stent ground truth (intra-observer performance: 75.5%) and 43.5% on all frames (including frames without stent, with guidewires, calcium, or without clinical use). The latter improved to 49.5% when gating the frames by the classification decision and further increased to 57.4% with a heuristic on the plausible stent strut area. Conclusions: A data-driven strategy for segmenting stents in ungated, manual pullbacks was presented—the most common and practical scenario in the time-critical clinical workflow. We demonstrated a mitigated risk for ambiguities and false positive predictions.
- 10 September 2021
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 118 (37)
Optimal ligand discrimination by asymmetric dimerization and turnover of interferon receptors
Significance: Upon viral infection, cytokine signaling molecules coordinate the immune response, but their distributions are highly inhomogeneous due to localized production and degradation in the infected tissue. It is a long-standing question how cytokine receptors can nevertheless decode inflammatory signals reliably. Here, we formulate this question as an inference problem in information theory and show that the molecular architecture of a receptor system determines its ability to discriminate between ligand types. Our results suggest that the type I interferon cytokine receptors have evolved the optimal design to detect and separate the presence of different ligand types in a noisy environment.
- 10 September 2021
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 118 (37)
Optimal ligand discrimination by asymmetric dimerization and turnover of interferon receptors
Significance: Upon viral infection, cytokine signaling molecules coordinate the immune response, but their distributions are highly inhomogeneous due to localized production and degradation in the infected tissue. It is a long-standing question how cytokine receptors can nevertheless decode inflammatory signals reliably. Here, we formulate this question as an inference problem in information theory and show that the molecular architecture of a receptor system determines its ability to discriminate between ligand types. Our results suggest that the type I interferon cytokine receptors have evolved the optimal design to detect and separate the presence of different ligand types in a noisy environment.