Multi-planar 2D-U-Net Segmentation of 3D-CT Abdominal Organs augmented by Spatial Occurrence Maps

Andre Mastmeyer; Daria Kern; Negar Chabi; Souraj Adhikary

arxiv: 2606.07717 · v1 · pith:VQJZYNQSnew · submitted 2026-06-05 · 📡 eess.IV · cs.AI· cs.CV

Multi-planar 2D-U-Net Segmentation of 3D-CT Abdominal Organs augmented by Spatial Occurrence Maps

Daria Kern , Negar Chabi , Souraj Adhikary , Andre Mastmeyer This is my paper

Pith reviewed 2026-06-27 20:13 UTC · model grok-4.3

classification 📡 eess.IV cs.AIcs.CV

keywords 2D U-NetCT organ segmentationspatial occurrence mapsmulti-planar segmentationabdominal organsDice scoremedical image segmentation

0 comments

The pith

Spatial occurrence maps improve multi-planar 2D U-Net Dice scores for abdominal organ segmentation by up to 4%.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to show that adding fuzzy 3D spatial occurrence maps to a multi-planar 2D U-Net architecture enhances segmentation of five abdominal organs in large field-of-view CT scans. A two-stage process first detects the volume of interest by axial traversal, then refines predictions inside those bounds using the maps as location cues. A sympathetic reader would care because the method promises higher accuracy while remaining computationally lighter than full 3D networks. Evaluation on 80 scans from public sources reports maximum Dice gains of about 4% over the identical model trained without the maps.

Core claim

The central claim is that augmenting multi-planar 2D-U-Net models with spatial occurrence maps supplies useful anatomical location cues, which improves segmentation accuracy for five abdominal organs in 3D CT scans and produces Dice score gains reaching about 4% compared to the baseline without these maps.

What carries the argument

Spatial occurrence maps, which are fuzzy 3D priors that encode anatomical location cues and augment the multi-planar 2D-U-Net inside the coarsely detected volume bounds.

Load-bearing premise

The spatial occurrence maps supply reliable anatomical location cues that remain useful and do not introduce bias when applied to new scans.

What would settle it

Apply the same trained models to an independent collection of CT scans acquired on different scanners or patient groups and measure whether the Dice improvement over the unaugmented baseline disappears.

Figures

Figures reproduced from arXiv: 2606.07717 by Andre Mastmeyer, Daria Kern, Negar Chabi, Souraj Adhikary.

**Figure 2.** Figure 2: Overview of the steps performed to extract the abdominal region. Left: Sagittal view of patient. Middle: [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Overview of fuzzy membership function, spatial map, occurrence map and the combined spatial occur [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: Organ group ground truth left to right: liver=green, spleen=light orange, right kidney=brown, left [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: Liver=green. Predicted segmentations=half-transparent yellow. (a) Peripheral surface represents well fit. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

**Figure 6.** Figure 6: Spleen=orange. Predicted segmentation=half-transparent yellow. (a) Peripheral area seems a good fit. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

**Figure 7.** Figure 7: Right kidney=brown. Predicted segmentations=half-transparent yellow. (a) The kidney is well fit. (b) [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

**Figure 8.** Figure 8: Left kidney=blue. Predicted segmentations=half-transparent yellow. (a) The kidney fit also reproduces a [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗

**Figure 9.** Figure 9: Pancreas=red. Predicted segmentations=half-transparent yellow. (a) The segmentation errors of this [PITH_FULL_IMAGE:figures/full_fig_p008_9.png] view at source ↗

read the original abstract

This work proposes a lightweight 2D-U-Net-based framework for segmenting five abdominal organs in large field-of-view 3D CT scans. The method combines coarse-to-fine segmentation, predictions from multiple anatomical planes, and additional fuzzy 3D spatial maps that provide anatomical location cues to improve segmentation accuracy. We combine multi-planar 2D-U-Net models augmented by a spatial occurrence map. The approach involves two main stages. First, the abdominal volume of interest region is detected by traversing the whole scan axially with a 2D-U-Net and determining the x-y-z-minimum and -maximum extents of the 5 abdominal organs of interest. Second, we use spatial occurrence maps to enhance our multi-planar 2D-U-net architecture inside the bounds from the former stage. The method is evaluated on 80 CT scans from various public sources. The results show Dice improvements of about 4% at maximum compared to the same model trained without spatial occurrence maps.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper reports a modest ~4% Dice gain from adding spatial occurrence maps to multi-planar 2D U-Nets on 80 CT scans, but provides almost no detail on map construction or evaluation protocol.

read the letter

The one thing to know is that this is an incremental methods paper showing a small Dice improvement from spatial priors in a 2D multi-planar setup for abdominal CT organ segmentation. The gain is presented as empirical and the framework stays lightweight, which is the main practical point.

What is new is the specific combination of coarse-to-fine VOI detection, multi-planar 2D U-Nets, and fuzzy 3D spatial occurrence maps for five organs. The two-stage pipeline makes sense for keeping compute low while giving the network location cues that pure image-based models might miss.

It does a few things cleanly: it works with public data sources, avoids full 3D convolutions, and reports a direct comparison to the same architecture without the maps. That comparison is the core result.

The soft spots are in the missing experimental details. The abstract gives no information on whether the spatial maps are computed only from training scans, how the 80-scan collection is split, or whether any cross-validation was used. If the maps incorporate statistics from the full set, the reported gain could be inflated by dataset-specific bias rather than general anatomical information. The phrase "about 4% at maximum" is also vague on which organs and what the variance looks like.

This is for people who build or tune segmentation pipelines in medical imaging and want to stay with 2D networks. A reader already working on similar U-Net variants would pick up the map idea and the VOI stage without much trouble.

It deserves peer review. The central claim is straightforward enough to test, and the work is honest about being an empirical tweak rather than a theoretical advance. Reviewers will likely press on the map construction and the evaluation protocol, which is exactly what should happen.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a lightweight 2D U-Net framework for segmenting five abdominal organs from large-FOV 3D CT scans. It employs a coarse-to-fine pipeline that first detects the abdominal volume of interest via axial traversal with a 2D U-Net, then performs multi-planar 2D U-Net segmentation inside those bounds, augmented by fuzzy 3D spatial occurrence maps that supply anatomical location cues. The method is evaluated on 80 CT scans drawn from various public sources and reports Dice-score improvements of up to approximately 4 % relative to the identical architecture trained without the spatial maps.

Significance. If the spatial occurrence maps are constructed exclusively from training data and the evaluation protocol is sound, the work would demonstrate a simple, computationally light mechanism for injecting anatomical priors into multi-planar 2D segmentation of abdominal CT. The combination of coarse localization, multi-planar inference, and occurrence-map augmentation is a plausible route to improved accuracy without resorting to full 3D networks; however, the absence of any description of map construction or experimental controls prevents assessment of whether the reported gain is reproducible or generalizable.

major comments (3)

[Abstract] Abstract (and, by extension, the Methods section): the central empirical claim is a maximum 4 % Dice improvement attributable to the spatial occurrence maps, yet the manuscript supplies no description of how these maps are generated, whether they are computed solely from the training subset of the 80-scan collection, or whether any test-scan statistics enter their construction. This information is load-bearing for the validity of the reported gain.
[Evaluation] Evaluation protocol (presumably §4 or §5): no information is given on the train/test partitioning, cross-validation scheme, or statistical testing used to establish the 4 % Dice improvement. With only 80 scans from multiple sources, the lack of these details leaves open the possibility that the observed difference reflects dataset-specific bias or overfitting rather than a generalizable anatomical cue.
[Methods] Methods (map-augmentation stage): the paper does not specify whether the spatial occurrence maps are fixed dataset-wide priors or are recomputed per training fold, nor does it report any ablation that isolates the contribution of the maps from the multi-planar or coarse-to-fine components. Without such controls the attribution of the Dice gain remains ambiguous.

minor comments (2)

[Abstract] Abstract: inconsistent capitalization (“2D-U-Net” versus “2D-U-net”) and the phrase “about 4% at maximum” would benefit from a more precise statement of the observed range across organs or folds.
[Introduction] The manuscript would be strengthened by explicit citation of prior work on spatial priors or occurrence maps in abdominal CT segmentation to clarify the incremental contribution.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments highlighting key omissions in our manuscript. We address each major point below and will revise the paper to supply the missing details on map construction, evaluation protocol, and experimental controls.

read point-by-point responses

Referee: [Abstract] Abstract (and, by extension, the Methods section): the central empirical claim is a maximum 4 % Dice improvement attributable to the spatial occurrence maps, yet the manuscript supplies no description of how these maps are generated, whether they are computed solely from the training subset of the 80-scan collection, or whether any test-scan statistics enter their construction. This information is load-bearing for the validity of the reported gain.

Authors: We agree that a description of map generation is absent from the current manuscript. The maps are constructed exclusively from the training subset by computing normalized voxel-wise occurrence frequencies of each organ from the training segmentations only; no test-scan information is used at any stage. We will add a dedicated subsection in Methods that fully specifies the construction procedure, including the exact normalization and the training-only constraint. revision: yes
Referee: [Evaluation] Evaluation protocol (presumably §4 or §5): no information is given on the train/test partitioning, cross-validation scheme, or statistical testing used to establish the 4 % Dice improvement. With only 80 scans from multiple sources, the lack of these details leaves open the possibility that the observed difference reflects dataset-specific bias or overfitting rather than a generalizable anatomical cue.

Authors: The manuscript indeed omits these protocol details. We will revise the Evaluation section to state the train/test partitioning scheme employed, confirm whether a single split or cross-validation was used, and report any statistical testing (e.g., paired tests on Dice scores) that supports the observed improvement. This will allow readers to assess reproducibility and generalizability. revision: yes
Referee: [Methods] Methods (map-augmentation stage): the paper does not specify whether the spatial occurrence maps are fixed dataset-wide priors or are recomputed per training fold, nor does it report any ablation that isolates the contribution of the maps from the multi-planar or coarse-to-fine components. Without such controls the attribution of the Dice gain remains ambiguous.

Authors: The maps are fixed dataset-wide priors computed once from the training data and are not recomputed per fold. The reported comparison holds the multi-planar and coarse-to-fine stages constant while toggling only the maps, thereby providing a direct control for their contribution. No additional component-wise ablations were performed. We will clarify the fixed-prior nature and the controlled comparison in the revised Methods; space permitting, we will also discuss whether further ablations can be included. revision: partial

Circularity Check

0 steps flagged

No significant circularity; purely empirical comparison

full rationale

The paper presents a multi-planar 2D U-Net segmentation pipeline augmented by spatial occurrence maps and reports an empirical Dice improvement of ~4% on 80 CT scans from public sources. No equations, derivations, fitted parameters renamed as predictions, or self-citation chains appear in the abstract or described method. The central result is a direct experimental comparison (with vs. without maps) rather than any quantity that reduces to its own inputs by construction. The evaluation protocol details are not provided here, but the absence of any mathematical derivation chain means no circularity of the enumerated kinds can be exhibited.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract-only review; the method implicitly assumes standard U-Net training behavior and that the introduced spatial maps act as useful priors.

axioms (1)

domain assumption Spatial occurrence maps derived from training data or anatomical knowledge provide unbiased location cues that generalize to new scans.
Central to the claimed improvement; invoked in the second stage of the pipeline.

invented entities (1)

Spatial occurrence maps no independent evidence
purpose: Fuzzy 3D anatomical location cues supplied as additional input to the U-Net.
Introduced to augment the multi-planar models; no independent evidence of correctness outside the reported Dice scores.

pith-pipeline@v0.9.1-grok · 5719 in / 1224 out tokens · 23211 ms · 2026-06-27T20:13:43.138597+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

26 extracted references · 2 canonical work pages

[1]

Medical Image Computing and Computer-Assisted Intervention , series =

Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas , keywords =. Medical Image Computing and Computer-Assisted Intervention , series =. 2015 , copyright =

2015
[2]

2016 , publisher =

Medical Image Computing and Computer-Assisted Intervention , series =. 2016 , publisher =

2016
[3]

2021 , journal=

Zettler, Nico and Mastmeyer, Andre , title =. 2021 , journal=

2021
[4]

Journal of Image and Graphics - JOIG , pages=

3D Bounding Box Detection in Volumetric Medical Image Data: A Systematic Literature Review , author=. Journal of Image and Graphics - JOIG , pages=. 2022 , note=

2022
[5]

3D U-Net abdominal organ segmentation in CT data using organ bounds , author=

2D vs. 3D U-Net abdominal organ segmentation in CT data using organ bounds , author=. Proc. SPIE Medical Imaging , number=
[6]

Nature methods , volume=

nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation , author=. Nature methods , volume=. 2021 , publisher=

2021
[7]

(https://arxiv.org/abs/1704.06382) , year=

Hierarchical 3D fully convolutional networks for multi-organ segmentation , author=. (https://arxiv.org/abs/1704.06382) , year=

Pith/arXiv arXiv
[8]

Medical Imaging 2020: Image-Guided Procedures, Robotic Interventions, and Modeling , volume=

CNN-based hierarchical coarse-to-fine segmentation of pelvic CT images for prostate cancer radiotherapy , author=. Medical Imaging 2020: Image-Guided Procedures, Robotic Interventions, and Modeling , volume=. 2020 , organization=

2020
[9]

IEEE Transactions on Medical Imaging , volume=

H-DenseUNet: hybrid densely connected UNet for liver and tumor segmentation from CT volumes , author=. IEEE Transactions on Medical Imaging , volume=. 2018 , publisher=

2018
[10]

Medical Image Computing and Computer-Assisted Intervention , pages=

Bridging the gap between 2d and 3d organ segmentation with volumetric fusion net , author=. Medical Image Computing and Computer-Assisted Intervention , pages=. 2018 , organization=

2018
[11]

International Journal of Computer Assisted Radiology and Surgery , volume=

Multi-dimensional consistency learning between 2D Swin U-Net and 3D U-Net for intestine segmentation from CT volume , author=. International Journal of Computer Assisted Radiology and Surgery , volume=. 2025 , publisher=

2025
[12]

iRADIOLOGY , volume =

Li, Qing and Zhang, Yizhe and Sun, Longyu and Sun, Mengting and Liu, Meng and Wang, Zian and Wang, Qi and Wang, Shuo and Wang, Chengyan , title =. iRADIOLOGY , volume =. doi:https://doi.org/10.1002/ird3.101 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1002/ird3.101 , abstract =

work page doi:10.1002/ird3.101
[13]

2023 , eprint=

Attention Is All You Need , author=. 2023 , eprint=

2023
[14]

Nature reviews neuroscience , volume=

Computational modelling of visual attention , author=. Nature reviews neuroscience , volume=. 2001 , publisher=

2001
[15]

Radiology: Artificial Intelligence , volume=

TotalSegmentator: robust segmentation of 104 anatomic structures in CT images , author=. Radiology: Artificial Intelligence , volume=. 2023 , publisher=

2023
[16]

Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

VISTA3D: A unified segmentation foundation model for 3D medical imaging , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=
[17]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Towards a comprehensive, efficient and promptable anatomic structure segmentation model using 3d whole-body ct scans , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=
[18]

arXiv preprint arXiv:2211.09562 , year=

Convolutional neural networks for medical image segmentation , author=. arXiv preprint arXiv:2211.09562 , year=

arXiv
[19]

AIP Conference Proceedings , volume=

A comprehensive review on CNN-based applications for medical imaging classification and segmentation , author=. AIP Conference Proceedings , volume=. 2024 , organization=

2024
[20]

Information , volume=

Deep convolutional neural networks in medical image analysis: A review , author=. Information , volume=. 2025 , publisher=

2025
[21]

Tomography , volume=

Medical image segmentation: A comprehensive review of deep learning-based methods , author=. Tomography , volume=. 2025 , publisher=

2025
[22]

German Conference on Medical Image Computing - BVM , pages=

Ray-casting-based evaluation framework for needle insertion force feedback algorithms , author=. German Conference on Medical Image Computing - BVM , pages=. 2013 , organization=

2013
[23]

Studies in Health Technology and Informatics , volume=

Optimized image-based soft tissue deformation algorithms for visualization of haptic needle insertion , author=. Studies in Health Technology and Informatics , volume=. 2013 , publisher=

2013
[24]

and Vosburgh, Kirby G

Kikinis, Ron and Pieper, Steve D. and Vosburgh, Kirby G. 3D Slicer: A Platform for Subject-Specific Image Analysis, Visualization, and Clinical Support. Intraoperative Imaging and Image-Guided Therapy. 2014. doi:10.1007/978-1-4614-7657-3_19

work page doi:10.1007/978-1-4614-7657-3_19 2014
[25]

Yushkevich and Joseph Piven and Cody Hazlett, Heather and Gimpel Smith, Rachel and Sean Ho and James C

Paul A. Yushkevich and Joseph Piven and Cody Hazlett, Heather and Gimpel Smith, Rachel and Sean Ho and James C. Gee and Guido Gerig , title =. Neuroimage , year =
[26]

BMC Medical Imaging , year=

From diverse CT scans to generalization: towards robust abdominal organ segmentation , author=. BMC Medical Imaging , year=

[1] [1]

Medical Image Computing and Computer-Assisted Intervention , series =

Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas , keywords =. Medical Image Computing and Computer-Assisted Intervention , series =. 2015 , copyright =

2015

[2] [2]

2016 , publisher =

Medical Image Computing and Computer-Assisted Intervention , series =. 2016 , publisher =

2016

[3] [3]

2021 , journal=

Zettler, Nico and Mastmeyer, Andre , title =. 2021 , journal=

2021

[4] [4]

Journal of Image and Graphics - JOIG , pages=

3D Bounding Box Detection in Volumetric Medical Image Data: A Systematic Literature Review , author=. Journal of Image and Graphics - JOIG , pages=. 2022 , note=

2022

[5] [5]

3D U-Net abdominal organ segmentation in CT data using organ bounds , author=

2D vs. 3D U-Net abdominal organ segmentation in CT data using organ bounds , author=. Proc. SPIE Medical Imaging , number=

[6] [6]

Nature methods , volume=

nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation , author=. Nature methods , volume=. 2021 , publisher=

2021

[7] [7]

(https://arxiv.org/abs/1704.06382) , year=

Hierarchical 3D fully convolutional networks for multi-organ segmentation , author=. (https://arxiv.org/abs/1704.06382) , year=

Pith/arXiv arXiv

[8] [8]

Medical Imaging 2020: Image-Guided Procedures, Robotic Interventions, and Modeling , volume=

CNN-based hierarchical coarse-to-fine segmentation of pelvic CT images for prostate cancer radiotherapy , author=. Medical Imaging 2020: Image-Guided Procedures, Robotic Interventions, and Modeling , volume=. 2020 , organization=

2020

[9] [9]

IEEE Transactions on Medical Imaging , volume=

H-DenseUNet: hybrid densely connected UNet for liver and tumor segmentation from CT volumes , author=. IEEE Transactions on Medical Imaging , volume=. 2018 , publisher=

2018

[10] [10]

Medical Image Computing and Computer-Assisted Intervention , pages=

Bridging the gap between 2d and 3d organ segmentation with volumetric fusion net , author=. Medical Image Computing and Computer-Assisted Intervention , pages=. 2018 , organization=

2018

[11] [11]

International Journal of Computer Assisted Radiology and Surgery , volume=

Multi-dimensional consistency learning between 2D Swin U-Net and 3D U-Net for intestine segmentation from CT volume , author=. International Journal of Computer Assisted Radiology and Surgery , volume=. 2025 , publisher=

2025

[12] [12]

iRADIOLOGY , volume =

Li, Qing and Zhang, Yizhe and Sun, Longyu and Sun, Mengting and Liu, Meng and Wang, Zian and Wang, Qi and Wang, Shuo and Wang, Chengyan , title =. iRADIOLOGY , volume =. doi:https://doi.org/10.1002/ird3.101 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1002/ird3.101 , abstract =

work page doi:10.1002/ird3.101

[13] [13]

2023 , eprint=

Attention Is All You Need , author=. 2023 , eprint=

2023

[14] [14]

Nature reviews neuroscience , volume=

Computational modelling of visual attention , author=. Nature reviews neuroscience , volume=. 2001 , publisher=

2001

[15] [15]

Radiology: Artificial Intelligence , volume=

TotalSegmentator: robust segmentation of 104 anatomic structures in CT images , author=. Radiology: Artificial Intelligence , volume=. 2023 , publisher=

2023

[16] [16]

Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

VISTA3D: A unified segmentation foundation model for 3D medical imaging , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

[17] [17]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Towards a comprehensive, efficient and promptable anatomic structure segmentation model using 3d whole-body ct scans , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

[18] [18]

arXiv preprint arXiv:2211.09562 , year=

Convolutional neural networks for medical image segmentation , author=. arXiv preprint arXiv:2211.09562 , year=

arXiv

[19] [19]

AIP Conference Proceedings , volume=

A comprehensive review on CNN-based applications for medical imaging classification and segmentation , author=. AIP Conference Proceedings , volume=. 2024 , organization=

2024

[20] [20]

Information , volume=

Deep convolutional neural networks in medical image analysis: A review , author=. Information , volume=. 2025 , publisher=

2025

[21] [21]

Tomography , volume=

Medical image segmentation: A comprehensive review of deep learning-based methods , author=. Tomography , volume=. 2025 , publisher=

2025

[22] [22]

German Conference on Medical Image Computing - BVM , pages=

Ray-casting-based evaluation framework for needle insertion force feedback algorithms , author=. German Conference on Medical Image Computing - BVM , pages=. 2013 , organization=

2013

[23] [23]

Studies in Health Technology and Informatics , volume=

Optimized image-based soft tissue deformation algorithms for visualization of haptic needle insertion , author=. Studies in Health Technology and Informatics , volume=. 2013 , publisher=

2013

[24] [24]

and Vosburgh, Kirby G

Kikinis, Ron and Pieper, Steve D. and Vosburgh, Kirby G. 3D Slicer: A Platform for Subject-Specific Image Analysis, Visualization, and Clinical Support. Intraoperative Imaging and Image-Guided Therapy. 2014. doi:10.1007/978-1-4614-7657-3_19

work page doi:10.1007/978-1-4614-7657-3_19 2014

[25] [25]

Yushkevich and Joseph Piven and Cody Hazlett, Heather and Gimpel Smith, Rachel and Sean Ho and James C

Paul A. Yushkevich and Joseph Piven and Cody Hazlett, Heather and Gimpel Smith, Rachel and Sean Ho and James C. Gee and Guido Gerig , title =. Neuroimage , year =

[26] [26]

BMC Medical Imaging , year=

From diverse CT scans to generalization: towards robust abdominal organ segmentation , author=. BMC Medical Imaging , year=