Duke Spark - Medical Imaging AI
@dukespark.bsky.social
Duke AI Health Initiative for Medical Imaging. We gather technical and clinical experts at Duke University to develop algorithms for analysis of medical images.
6/6
CSpineSeg is published on Medical Imaging and Data Resource Center, data.midrc.org/discovery/H6.... The full data descriptor has just been published with Springer Nature in Scientific Data. Read here: rdcu.be/eMWgW.
CSpineSeg is published on Medical Imaging and Data Resource Center, data.midrc.org/discovery/H6.... The full data descriptor has just been published with Springer Nature in Scientific Data. Read here: rdcu.be/eMWgW.
November 7, 2025 at 3:50 PM
6/6
CSpineSeg is published on Medical Imaging and Data Resource Center, data.midrc.org/discovery/H6.... The full data descriptor has just been published with Springer Nature in Scientific Data. Read here: rdcu.be/eMWgW.
CSpineSeg is published on Medical Imaging and Data Resource Center, data.midrc.org/discovery/H6.... The full data descriptor has just been published with Springer Nature in Scientific Data. Read here: rdcu.be/eMWgW.
5/6
CSpineSeg can serve as a valuable resource for the development and validation of deep learning models, as well as for clinical investigators focusing on the diagnosis of cervical spine diseases.
CSpineSeg can serve as a valuable resource for the development and validation of deep learning models, as well as for clinical investigators focusing on the diagnosis of cervical spine diseases.
November 7, 2025 at 3:50 PM
5/6
CSpineSeg can serve as a valuable resource for the development and validation of deep learning models, as well as for clinical investigators focusing on the diagnosis of cervical spine diseases.
CSpineSeg can serve as a valuable resource for the development and validation of deep learning models, as well as for clinical investigators focusing on the diagnosis of cervical spine diseases.
4/6
A deep-learning-based segmentation model was trained on labeled MRIs, internally validated with robust performance (DSC>90%), and subsequently automatically labeled unannotated MRIs.
A deep-learning-based segmentation model was trained on labeled MRIs, internally validated with robust performance (DSC>90%), and subsequently automatically labeled unannotated MRIs.
November 7, 2025 at 3:50 PM
4/6
A deep-learning-based segmentation model was trained on labeled MRIs, internally validated with robust performance (DSC>90%), and subsequently automatically labeled unannotated MRIs.
A deep-learning-based segmentation model was trained on labeled MRIs, internally validated with robust performance (DSC>90%), and subsequently automatically labeled unannotated MRIs.
3/6
We invited 6 board-certified radiologists to annotate vertebral bodies and intervertebral discs on 491 MRIs (~40% of the data).
We invited 6 board-certified radiologists to annotate vertebral bodies and intervertebral discs on 491 MRIs (~40% of the data).
November 7, 2025 at 3:50 PM
3/6
We invited 6 board-certified radiologists to annotate vertebral bodies and intervertebral discs on 491 MRIs (~40% of the data).
We invited 6 board-certified radiologists to annotate vertebral bodies and intervertebral discs on 491 MRIs (~40% of the data).
2/6
CSpineSeg fills a crucial need for cervical spine imaging research on MRI by providing comprehensive vertebral body and intervertebral discs segmentation. It includes 1255 sagittal T2-weighted cervical spine MRIs collected from 1232 patients at Duke University Health System.
CSpineSeg fills a crucial need for cervical spine imaging research on MRI by providing comprehensive vertebral body and intervertebral discs segmentation. It includes 1255 sagittal T2-weighted cervical spine MRIs collected from 1232 patients at Duke University Health System.
November 7, 2025 at 3:50 PM
2/6
CSpineSeg fills a crucial need for cervical spine imaging research on MRI by providing comprehensive vertebral body and intervertebral discs segmentation. It includes 1255 sagittal T2-weighted cervical spine MRIs collected from 1232 patients at Duke University Health System.
CSpineSeg fills a crucial need for cervical spine imaging research on MRI by providing comprehensive vertebral body and intervertebral discs segmentation. It includes 1255 sagittal T2-weighted cervical spine MRIs collected from 1232 patients at Duke University Health System.
10/10
🔹Epidemiologic trends reproduced (e.g., density decreases with age), adding confidence that the MRI metric behaves as expected at population scale.
The full paper can be found at: arxiv.org/abs/2504.15192
The segmentation model can be found at: github.com/mazurowski-l...
🔹Epidemiologic trends reproduced (e.g., density decreases with age), adding confidence that the MRI metric behaves as expected at population scale.
The full paper can be found at: arxiv.org/abs/2504.15192
The segmentation model can be found at: github.com/mazurowski-l...
October 31, 2025 at 2:48 PM
10/10
🔹Epidemiologic trends reproduced (e.g., density decreases with age), adding confidence that the MRI metric behaves as expected at population scale.
The full paper can be found at: arxiv.org/abs/2504.15192
The segmentation model can be found at: github.com/mazurowski-l...
🔹Epidemiologic trends reproduced (e.g., density decreases with age), adding confidence that the MRI metric behaves as expected at population scale.
The full paper can be found at: arxiv.org/abs/2504.15192
The segmentation model can be found at: github.com/mazurowski-l...
9/10
What we found:
🔹Strong alignment with mammography, while MRI’s volumetric view uncovers meaningful within-category heterogeneity, especially in very dense cases
🔹Robust generalization across all 3 datasets, suggesting readiness for multi-site research and longitudinal use
What we found:
🔹Strong alignment with mammography, while MRI’s volumetric view uncovers meaningful within-category heterogeneity, especially in very dense cases
🔹Robust generalization across all 3 datasets, suggesting readiness for multi-site research and longitudinal use
October 31, 2025 at 2:48 PM
9/10
What we found:
🔹Strong alignment with mammography, while MRI’s volumetric view uncovers meaningful within-category heterogeneity, especially in very dense cases
🔹Robust generalization across all 3 datasets, suggesting readiness for multi-site research and longitudinal use
What we found:
🔹Strong alignment with mammography, while MRI’s volumetric view uncovers meaningful within-category heterogeneity, especially in very dense cases
🔹Robust generalization across all 3 datasets, suggesting readiness for multi-site research and longitudinal use
8/10
🔹Transparent, modular design so future improvements to components (e.g., localization, normalization) won’t break comparability of the final density metric.
🔹Transparent, modular design so future improvements to components (e.g., localization, normalization) won’t break comparability of the final density metric.
October 31, 2025 at 2:48 PM
8/10
🔹Transparent, modular design so future improvements to components (e.g., localization, normalization) won’t break comparability of the final density metric.
🔹Transparent, modular design so future improvements to components (e.g., localization, normalization) won’t break comparability of the final density metric.
7/10
🔹A head-to-head comparison framework aligning MRI-based density with mammographic categories to assess concordance and reveal within-category variation.
🔹A head-to-head comparison framework aligning MRI-based density with mammographic categories to assess concordance and reveal within-category variation.
October 31, 2025 at 2:48 PM
7/10
🔹A head-to-head comparison framework aligning MRI-based density with mammographic categories to assess concordance and reveal within-category variation.
🔹A head-to-head comparison framework aligning MRI-based density with mammographic categories to assess concordance and reveal within-category variation.
6/10
What we built:
🔹A standardized, end-to-end MRI pipeline for volumetric density (from breast/FGT localization to final density metric), designed to be scalable and protocol-agnostic.
What we built:
🔹A standardized, end-to-end MRI pipeline for volumetric density (from breast/FGT localization to final density metric), designed to be scalable and protocol-agnostic.
October 31, 2025 at 2:48 PM
6/10
What we built:
🔹A standardized, end-to-end MRI pipeline for volumetric density (from breast/FGT localization to final density metric), designed to be scalable and protocol-agnostic.
What we built:
🔹A standardized, end-to-end MRI pipeline for volumetric density (from breast/FGT localization to final density metric), designed to be scalable and protocol-agnostic.
5/10
Three independent datasets, real-world rigor:
We validated across three distinct cohorts -screening, clinical (heterogeneous, real-world MRIs), and an external dataset- covering different scanners, protocols, and populations (nearly 2,000 MRIs in total).
Three independent datasets, real-world rigor:
We validated across three distinct cohorts -screening, clinical (heterogeneous, real-world MRIs), and an external dataset- covering different scanners, protocols, and populations (nearly 2,000 MRIs in total).
October 31, 2025 at 2:48 PM
5/10
Three independent datasets, real-world rigor:
We validated across three distinct cohorts -screening, clinical (heterogeneous, real-world MRIs), and an external dataset- covering different scanners, protocols, and populations (nearly 2,000 MRIs in total).
Three independent datasets, real-world rigor:
We validated across three distinct cohorts -screening, clinical (heterogeneous, real-world MRIs), and an external dataset- covering different scanners, protocols, and populations (nearly 2,000 MRIs in total).
4/10
We built a unified MRI workflow so density can be measured consistently across sites and studies, and then we directly compared MRI-derived density with mammographic density to see where they agree and where MRI adds nuance
We built a unified MRI workflow so density can be measured consistently across sites and studies, and then we directly compared MRI-derived density with mammographic density to see where they agree and where MRI adds nuance
October 31, 2025 at 2:48 PM
4/10
We built a unified MRI workflow so density can be measured consistently across sites and studies, and then we directly compared MRI-derived density with mammographic density to see where they agree and where MRI adds nuance
We built a unified MRI workflow so density can be measured consistently across sites and studies, and then we directly compared MRI-derived density with mammographic density to see where they agree and where MRI adds nuance
3/10
Why these matters:
Mammography gives widely used density categories, but MRI offers true 3D anatomy
Why these matters:
Mammography gives widely used density categories, but MRI offers true 3D anatomy
October 31, 2025 at 2:48 PM
3/10
Why these matters:
Mammography gives widely used density categories, but MRI offers true 3D anatomy
Why these matters:
Mammography gives widely used density categories, but MRI offers true 3D anatomy
2/10
We are excited to share our work introducing a standard, end-to-end pipeline to compute breast density directly from MRI, a reproducible alternative that complements traditional mammography-based density
We are excited to share our work introducing a standard, end-to-end pipeline to compute breast density directly from MRI, a reproducible alternative that complements traditional mammography-based density
October 31, 2025 at 2:48 PM
2/10
We are excited to share our work introducing a standard, end-to-end pipeline to compute breast density directly from MRI, a reproducible alternative that complements traditional mammography-based density
We are excited to share our work introducing a standard, end-to-end pipeline to compute breast density directly from MRI, a reproducible alternative that complements traditional mammography-based density
7/7
🔹Composable with other TTA methods: stacks on approaches like SITA/FSeg for further boosts
Stable, label-free adaptation from just one image—for realistic medical workflows.
The full paper can be found at openaccess.thecvf.com/content/CVPR...
🔹Composable with other TTA methods: stacks on approaches like SITA/FSeg for further boosts
Stable, label-free adaptation from just one image—for realistic medical workflows.
The full paper can be found at openaccess.thecvf.com/content/CVPR...
October 24, 2025 at 2:11 PM
7/7
🔹Composable with other TTA methods: stacks on approaches like SITA/FSeg for further boosts
Stable, label-free adaptation from just one image—for realistic medical workflows.
The full paper can be found at openaccess.thecvf.com/content/CVPR...
🔹Composable with other TTA methods: stacks on approaches like SITA/FSeg for further boosts
Stable, label-free adaptation from just one image—for realistic medical workflows.
The full paper can be found at openaccess.thecvf.com/content/CVPR...
6/7
🔹Structure-aware confidence: balanced foreground/background entropy weighting yields sharper boundaries and fewer over-confident false positives
🔹BN-statistics done right: robustness by integrating over train/test/mixed statistics instead of betting on one unstable choice
🔹Structure-aware confidence: balanced foreground/background entropy weighting yields sharper boundaries and fewer over-confident false positives
🔹BN-statistics done right: robustness by integrating over train/test/mixed statistics instead of betting on one unstable choice
October 24, 2025 at 2:11 PM
6/7
🔹Structure-aware confidence: balanced foreground/background entropy weighting yields sharper boundaries and fewer over-confident false positives
🔹BN-statistics done right: robustness by integrating over train/test/mixed statistics instead of betting on one unstable choice
🔹Structure-aware confidence: balanced foreground/background entropy weighting yields sharper boundaries and fewer over-confident false positives
🔹BN-statistics done right: robustness by integrating over train/test/mixed statistics instead of betting on one unstable choice
5/7
With InTEnt, we achieve:
🔹Single-image TTA that works: no batches, no labels, no training-time changes—plug-and-play on a pretrained segmenter
🔹Consistent gains across shifts: +2.9% Dice on average vs. the leading method across 24 source/target splits and 3 datasets
With InTEnt, we achieve:
🔹Single-image TTA that works: no batches, no labels, no training-time changes—plug-and-play on a pretrained segmenter
🔹Consistent gains across shifts: +2.9% Dice on average vs. the leading method across 24 source/target splits and 3 datasets
October 24, 2025 at 2:11 PM
5/7
With InTEnt, we achieve:
🔹Single-image TTA that works: no batches, no labels, no training-time changes—plug-and-play on a pretrained segmenter
🔹Consistent gains across shifts: +2.9% Dice on average vs. the leading method across 24 source/target splits and 3 datasets
With InTEnt, we achieve:
🔹Single-image TTA that works: no batches, no labels, no training-time changes—plug-and-play on a pretrained segmenter
🔹Consistent gains across shifts: +2.9% Dice on average vs. the leading method across 24 source/target splits and 3 datasets
4/7
At its core is a simple but powerful idea:
integrate over predictions made with multiple BN statistic choices (from train → test and mixes in between), then weight those predictions by balanced entropy (foreground/background) to produce a stable, structure-preserving mask.
At its core is a simple but powerful idea:
integrate over predictions made with multiple BN statistic choices (from train → test and mixes in between), then weight those predictions by balanced entropy (foreground/background) to produce a stable, structure-preserving mask.
October 24, 2025 at 2:11 PM
4/7
At its core is a simple but powerful idea:
integrate over predictions made with multiple BN statistic choices (from train → test and mixes in between), then weight those predictions by balanced entropy (foreground/background) to produce a stable, structure-preserving mask.
At its core is a simple but powerful idea:
integrate over predictions made with multiple BN statistic choices (from train → test and mixes in between), then weight those predictions by balanced entropy (foreground/background) to produce a stable, structure-preserving mask.
3/7
We present InTEnt — Integrated Entropy Weighting for Single-Image Test-Time Adaptation. InTEnt adapts any BN-based segmentation model from a single test image without retraining or extra networks.
We present InTEnt — Integrated Entropy Weighting for Single-Image Test-Time Adaptation. InTEnt adapts any BN-based segmentation model from a single test image without retraining or extra networks.
October 24, 2025 at 2:11 PM
3/7
We present InTEnt — Integrated Entropy Weighting for Single-Image Test-Time Adaptation. InTEnt adapts any BN-based segmentation model from a single test image without retraining or extra networks.
We present InTEnt — Integrated Entropy Weighting for Single-Image Test-Time Adaptation. InTEnt adapts any BN-based segmentation model from a single test image without retraining or extra networks.
2/7
But most TTA methods assume batches of target images - an unrealistic constraint in clinical settings - so performance often collapses when you only have one unlabeled test image.
But most TTA methods assume batches of target images - an unrealistic constraint in clinical settings - so performance often collapses when you only have one unlabeled test image.
October 24, 2025 at 2:11 PM
2/7
But most TTA methods assume batches of target images - an unrealistic constraint in clinical settings - so performance often collapses when you only have one unlabeled test image.
But most TTA methods assume batches of target images - an unrealistic constraint in clinical settings - so performance often collapses when you only have one unlabeled test image.
6/6
🔹Robustness across diverse datasets and imaging protocols
🔹Stable and reproducible translation performance at scale
The full paper is available at arxiv.org/abs/2403.107..., and the ContourDiff codebase can be accessed at github.com/mazurowski-l...
🔹Robustness across diverse datasets and imaging protocols
🔹Stable and reproducible translation performance at scale
The full paper is available at arxiv.org/abs/2403.107..., and the ContourDiff codebase can be accessed at github.com/mazurowski-l...
October 17, 2025 at 2:26 PM
6/6
🔹Robustness across diverse datasets and imaging protocols
🔹Stable and reproducible translation performance at scale
The full paper is available at arxiv.org/abs/2403.107..., and the ContourDiff codebase can be accessed at github.com/mazurowski-l...
🔹Robustness across diverse datasets and imaging protocols
🔹Stable and reproducible translation performance at scale
The full paper is available at arxiv.org/abs/2403.107..., and the ContourDiff codebase can be accessed at github.com/mazurowski-l...