pith. sign in

arxiv: 1906.10338 · v1 · pith:XOR6NLJFnew · submitted 2019-06-25 · 📡 eess.IV · cs.CV· cs.LG· physics.med-ph

Learning a sparse database for patch-based medical image segmentation

Moti Freiman , Hannes Nickisch , Holger Schmitt , Pal Maurovich-Horvat , Patrick Donnelly , Mani Vembar , Liran Goshen This is my paper

Pith reviewed 2026-05-25 16:39 UTC · model grok-4.3

classification 📡 eess.IV cs.CVcs.LGphysics.med-ph
keywords patch-based segmentationdatabase optimizationcoronary lumen segmentationCCTAenergy minimizationsparse databasefractional flow reserve
0
0 comments X

The pith

Optimizing a database for patch-based segmentation reduces its size by 96% while preserving accuracy and improving FFR specificity.

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

The paper introduces a functional for learning an optimal database for patch-based image segmentation, consisting of fidelity, sparseness, and robustness to small-variations terms. This turns the database optimization into an energy minimization problem that can be solved with standard numerical tools, unlike rule-based prototype selection. Applied to coronary lumen segmentation from CCTA data, the method produces a database 96% smaller that maintains segmentation accuracy. It also shows improved specificity in CCTA-based fractional flow reserve estimation using the optimized database.

Core claim

The central discovery is that minimizing the proposed functional on a training set yields a sparse database for patch-based coronary lumen segmentation that achieves the same accuracy as the full database but with only 4% of the entries, and provides higher specificity (0.73 versus 0.7) for fractional flow reserve computation on a set of 132 lesions.

What carries the argument

The energy functional with fidelity, sparseness and robustness terms minimized numerically to select the optimal sparse database.

If this is right

  • Database size can be reduced by 96% without loss in lumen segmentation accuracy.
  • Improved specificity for CCTA based fractional flow reserve is achieved (0.73 vs 0.7 for all lesions).
  • The optimization applies to the publicly available MICCAI 2012 data.
  • The method formulates database learning as a solvable energy minimization rather than heuristic rules.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar optimization could be used to create compact databases for other patch-based medical imaging tasks.
  • The robustness term may help the database generalize to variations in image acquisition not seen in training.
  • Smaller databases could reduce memory requirements and speed up segmentation in clinical software.

Load-bearing premise

The weights on the fidelity, sparseness, and robustness terms are such that the energy minimum on training data corresponds to optimal performance on new data.

What would settle it

Evaluating the optimized database on an independent set of CCTA images from a different center and observing that the segmentation accuracy falls below the level achieved by the original database.

Figures

Figures reproduced from arXiv: 1906.10338 by Hannes Nickisch, Holger Schmitt, Liran Goshen, Mani Vembar, Moti Freiman, Pal Maurovich-Horvat, Patrick Donnelly.

Figure 1
Figure 1. Figure 1: The database reduction for each bin of the histogram of pro [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Representative example of straight multi-planar reconstru [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
read the original abstract

We introduce a functional for the learning of an optimal database for patch-based image segmentation with application to coronary lumen segmentation from coronary computed tomography angiography (CCTA) data. The proposed functional consists of fidelity, sparseness and robustness to small-variations terms and their associated weights. Existing work address database optimization by prototype selection aiming to optimize the database by either adding or removing prototypes according to a set of predefined rules. In contrast, we formulate the database optimization task as an energy minimization problem that can be solved using standard numerical tools. We apply the proposed database optimization functional to the task of optimizing a database for patch-base coronary lumen segmentation. Our experiments using the publicly available MICCAI 2012 coronary lumen segmentation challenge data show that optimizing the database using the proposed approach reduced database size by 96% while maintaining the same level of lumen segmentation accuracy. Moreover, we show that the optimized database yields an improved specificity of CCTA based fractional flow reserve (0.73 vs 0.7 for all lesions and 0.68 vs 0.65 for obstructive lesions) using a training set of 132 (76 obstructive) coronary lesions with invasively measured FFR as the reference.

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.

Referee Report

1 major / 0 minor

Summary. The manuscript proposes an energy functional with fidelity, sparseness, and robustness terms (and associated weights) to formulate database optimization for patch-based segmentation as a numerical minimization problem. Applied to coronary lumen segmentation from CCTA, experiments on the public MICCAI 2012 challenge data report a 96% reduction in database size while preserving lumen segmentation accuracy, plus improved CCTA-based FFR specificity (0.73 vs. 0.7 overall; 0.68 vs. 0.65 for obstructive lesions) on a separate set of 132 lesions.

Significance. If the numerical claims are robust, the work supplies a principled, solver-compatible alternative to rule-based prototype selection for patch databases. The use of public challenge data and the downstream clinical metric (FFR specificity) are positive features; a 96% size reduction without accuracy loss would be practically useful for memory- and compute-constrained medical imaging pipelines.

major comments (1)
  1. [Abstract] Abstract (and results section): the central claims of 96% database reduction and FFR specificity gains (0.73/0.68 vs. 0.7/0.65) are presented without error bars, cross-validation protocol, statistical significance tests, or ablation experiments isolating the three energy terms. These omissions leave open whether the reported performance is governed by the energy minimum or by post-hoc weight selection on the specific datasets.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the positive evaluation of the significance of our work and for the constructive comments. We address the major comment below and will incorporate the requested additions in the revised manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract (and results section): the central claims of 96% database reduction and FFR specificity gains (0.73/0.68 vs. 0.7/0.65) are presented without error bars, cross-validation protocol, statistical significance tests, or ablation experiments isolating the three energy terms. These omissions leave open whether the reported performance is governed by the energy minimum or by post-hoc weight selection on the specific datasets.

    Authors: We agree that the current manuscript does not include error bars, an explicit description of the cross-validation protocol, statistical significance tests, or ablation experiments on the individual energy terms. In the revised version we will add these elements: error bars (standard deviations from the MICCAI 2012 cross-validation folds), a clear statement of the evaluation protocol for both the segmentation and FFR experiments, paired statistical tests (Wilcoxon signed-rank for accuracy and McNemar’s test for specificity), and ablation results obtained by minimizing the functional with each term removed in turn. The weights are not chosen post-hoc on test data; they are part of the joint numerical minimization performed on the training set. We will report the optimized weight values and the solver settings to make this explicit. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical validation on external data

full rationale

The paper defines an energy functional (fidelity + sparseness + robustness terms) independently of any target performance metric, formulates database optimization as standard numerical minimization of that functional, and reports measured outcomes (96% size reduction, maintained lumen accuracy, improved FFR specificity) on the public MICCAI 2012 challenge dataset. No derivation step equates a claimed result to its own inputs by construction, no fitted parameter is relabeled as a prediction, and no load-bearing premise rests on self-citation chains. The central claims are falsifiable empirical measurements on held-out external data rather than algebraic identities.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

The central claim rests on the premise that the three-term energy functional can be minimized to yield a database whose segmentation performance matches the full set; no explicit free parameters beyond the term weights are stated, and no new physical entities are introduced.

free parameters (1)
  • weights of fidelity, sparseness and robustness terms
    The functional is defined with associated weights whose specific values are not given in the abstract and must be chosen or fitted to achieve the reported size reduction.

pith-pipeline@v0.9.0 · 5771 in / 1208 out tokens · 28227 ms · 2026-05-25T16:39:23.544862+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

11 extracted references · 11 canonical work pages

  1. [1]

    Annals of Applied Statistics 5(4), 2403–2424 (2011)

    Bien, J., Tibshirani, R.: Prototype selection for interp retable classification. Annals of Applied Statistics 5(4), 2403–2424 (2011)

    work page 2011

  2. [2]

    IEEE Trans

    Cover, T., Hart, P.: Nearest neighbor pattern classificat ion. IEEE Trans. Inform. Theory 13(1), 21–27 (1967)

    work page 1967

  3. [3]

    Med Phys 44(3), 1040–1049 (2017)

    Freiman, M., et al.: Improving CCTA-based lesions’ hemod ynamic significance assessment by accounting for partial volume modeling in aut omatic coronary lumen segmentation. Med Phys 44(3), 1040–1049 (2017)

    work page 2017

  4. [4]

    IEEE Trans Patt Anal Mach Intell 3 4(3), 417–435 (2012)

    Garc ´ ıa, S., et al.: Prototype Selection for Nearest Neig hbor Classification: Tax- onomy and Empirical Study. IEEE Trans Patt Anal Mach Intell 3 4(3), 417–435 (2012)

    work page 2012

  5. [5]

    Med Image Anal 1 7(8), 859–876 (2013), http://coronary.bigr.nl/stenoses/

    Kiri¸ sli, H., et al.: Standardized evaluation framework for evaluating coronary artery stenosis detection, stenosis quantification and lum en segmentation algo- rithms in computed tomography angiography. Med Image Anal 1 7(8), 859–876 (2013), http://coronary.bigr.nl/stenoses/

    work page 2013

  6. [6]

    In: Menze, B., et al

    Lugauer, F., et al.: Precise Lumen Segmentation in Corona ry Computed Tomogra- phy Angiography. In: Menze, B., et al. (eds.) Proc. of MICCAI 2014 Workshop on Medical Computer Vision: Algorithms for Big Data, Lecture N otes in Computer Science, vol. 8848, pp. 137–147. Springer (2014)

    work page 2014

  7. [7]

    In t J Biomed Imaging 2016, Article ID 7952541 (2016)

    Mechrez, R., Goldberger, J., Greenspan, H.: Patch-Based Segmentation with Spa- tial Consistency: Application to MS Lesions in Brain MRI. In t J Biomed Imaging 2016, Article ID 7952541 (2016)

    work page 2016

  8. [8]

    In: Navab, N., et al

    Nickisch, H., et al.: Learning Patient-Specific Lumped Mo dels for Interactive Coro- nary Blood Flow Simulations. In: Navab, N., et al. (eds.) Med ical Image Computing and Computer-Assisted Intervention – MICCAI 2015, Lecture Notes in Computer Science, vol. 9350, pp. 433–441. Springer International Pu blishing (2015)

    work page 2015

  9. [9]

    IEEE Trans Med Imaging 24(4), 477–485 (2005)

    Olabarriaga, S.D., et al.: Segmentation of thrombus in ab dominal aortic aneurysms from CTA with nonparametric statistical grey level appeara nce modeling. IEEE Trans Med Imaging 24(4), 477–485 (2005)

    work page 2005

  10. [10]

    NA Report NA2009/06 p

    Powell, M.: The BOBYQA algorithm for bound constrained o ptimization with- out derivatives. NA Report NA2009/06 p. 39 (2009), http://www6.cityu.edu. hk/rcms/publications/preprint26.pdf

    work page 2009

  11. [11]

    In : Lecture Notes in Com- puter Science (including subseries Lecture Notes in Artific ial Intelligence and Lec- ture Notes in Bioinformatics)

    Wang, Z., et al.: Geodesic patch-based segmentation. In : Lecture Notes in Com- puter Science (including subseries Lecture Notes in Artific ial Intelligence and Lec- ture Notes in Bioinformatics). vol. 8673 LNCS, pp. 666–673 ( 2014)

    work page 2014