Abstract
Incorporating prior knowledge into a segmentation task, whether it be under the form of geometrical constraints (area/volume penalisation, convexity enforcement, etc.) or of topological constraints (to preserve the contextual relations between objects, to monitor the number of connected components), proves to increase accuracy in medical image segmentation. In particular, it allows to compensate for the issue of weak boundary definition, of imbalanced classes, and to be more in line with anatomical consistency even though the data do not explicitly exhibit those features. This observation underpins the introduced contribution that aims, in a hybrid setting, to leverage the best of both worlds that variational methods and supervised deep learning approaches embody: (a) versatility and adaptability in the mathematical formulation of the problem to encode geometrical/topological constraints, (b) interpretability of the results for the former formalism, while (c) more efficient and effective processing models, (d) ability to become more proficient at learning intricate features and executing more computationally intensive tasks, for the latter one. To be more precise, a unified variational framework involving topological prescriptions in the training of convolutional neural networks through the design of a suitable penalty in the loss function is provided. These topological constraints are implicitly enforced by viewing the segmentation procedure as a registration task between the processed image and its associated ground truth under incompressibility conditions, thus making them homeomorphic. A very preliminary version (Lambert et al., in Calatroni, Donatelli, Morigi, Prato, Santacesaria (eds) Scale space and variational methods in computer vision, Springer, Berlin, 2023, pp. 363–375) of this work has been published in the proceedings of the Ninth International Conference on Scale Space and Variational Methods in Computer Vision, 2023. It contained neither all the theoretical results, nor the detailed related proofs, nor did it include the numerical analysis of the designed algorithm. Besides these more involved developments in the present version, a more complete, systematic and thorough analysis of the numerical experiments is also conducted, addressing several issues: (i) limited amount of labelled data in the training phase, (ii) low contrast or imbalanced classes exhibited by the data, and (iii) explainability of the results.
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Fige_HTML.png)
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Figh_HTML.png)
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Fig1_HTML.png)
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Fig2_HTML.png)
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Fig3_HTML.png)
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Fig4_HTML.png)
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Fig5_HTML.png)
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Fig6_HTML.png)
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Fig7_HTML.jpg)
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Fig8_HTML.jpg)
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Fig9_HTML.png)
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Fig10_HTML.png)
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Fig11_HTML.png)
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Fig12_HTML.png)
![](http://media.springernature.com/m312/springer-static/image/art%3A10.1007%2Fs10851-024-01172-3/MediaObjects/10851_2024_1172_Fig13_HTML.png)
Similar content being viewed by others
Data Availability
The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request
References
Allaire, G.: Analyse Numérique et Optimisation. Les Éditions de l’École Polytechnique, Palaiseau (2007)
Antonelli, M., Reinke, A., Bakas, S., Farahani, K., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B., Ronneberger, O., Summers, R.M., et al.: The medical segmentation decathlon. Nat. Commun. 13(1), 4128 (2022)
Ball, J.M.: Global invertibility of Sobolev functions and the interpenetration of matter. P. R. Soc. Edinb. A 88(3–4), 315–328 (1981)
Boutry, N., Géraud, T., Najman, L.: A tutorial on well-composedness. J. Math. Imaging Vis. 60(3), 443–478 (2018)
Brezis, H.: Analyse fonctionnelle. Dunod, Paris (2005)
Ciarlet, P.: Three-Dimensional Elasticity. Mathematical Elasticity. Elsevier, Amsterdam (1994)
Clough, J., Byrne, N., Oksuz, I., Zimmer, V.A., Schnabel, J.A., King, A.: A topological loss function for deep-learning based image segmentation using persistent homology. IEEE Trans. Pattern Anal. Mach. Intell. 6, 66 (2020)
Dacorogna, B.: Direct Methods in the Calculus of Variations, 2nd edn. Springer, Berlin (2008)
Debroux, N., Aston, J., Bonardi, F., Forbes, A., Le Guyader, C., Romanchikova, M., Schönlieb, C.B.: A variational model dedicated to joint segmentation, registration, and atlas generation for shape analysis. SIAM J. Imaging Sci. 13(1), 351–380 (2020)
Debroux, N., Le Guyader, C.: A joint segmentation/registration model based on a nonlocal characterization of weighted total variation and nonlocal shape descriptors. SIAM J. Imaging Sci. 11(2), 957–990 (2018)
Demengel, F., Demengel, G., Erné, R.: Functional Spaces for the Theory of Elliptic Partial Differential Equations. Universitext, Springer, London (2012)
Edelsbrunner, H., Harer, J.L.: Computational Topology: An Introduction. Applied Mathematics, American Mathematical Society, Philadelphia (2010)
El Jurdi, R., Petitjean, C., Honeine, P., Cheplygina, V., Abdallah, F.: High-level prior-based loss functions for medical image segmentation: a survey. Comput. Vis. Image Underst. 210, 103248 (2021)
Estienne, T., Vakalopoulou, M., Christodoulidis, S., Battistela, E., Lerousseau, M., Carre, A., Klausner, G., Sun, R., Robert, C., Mougiakakou, S., Paragios, N., Deutsch, E.: U-ReSNet: ultimate coupling of registration and segmentation with deep nets. In: Shen, D., Liu, T., Peters, T.M., Staib, L.H., Essert, C, Zhou, S., Yap, P.T., Khan, A. (Eds.) Medical Image Computing and Computer Assisted Intervention—MICCAI 2019. Springer, pp. 310–319 (2019)
Evans, L., Gariepy, R.: Measure Theory and Fine Properties of Functions. Studies in Advanced Mathematics. Taylor & Francis, London (1991)
Glorot, X., Bengio, Y.: Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics. JMLR Workshop and Conference Proceedings, pp. 249–256 (2010)
Glowinski, R., Le Tallec, P.: Numerical solution of problems in incompressible finite elasticity by augmented Lagrangian methods. I. Two-dimensional and axisymmetric problems. SIAM J. Appl. Math. 42(2), 400–429 (1982)
Han, X., Xu, C., Braga-Neto, U., Prince, J.L.: Topology correction in brain cortex segmentation using a multiscale, graph-based algorithm. IEEE Trans. Med. Imaging 21(2), 109–121 (2002)
Han, X., Xu, C., Prince, J.L.: A topology preserving level set method for geometric deformable models. IEEE Trans. Pattern Anal. Mach. Intell. 25(6), 755–768 (2003)
Hu, X., Li, F., Samaras, D., Chen, C.: Topology-preserving deep image segmentation. Adv. Neural Inf. Process. Syst. 32, 66 (2019)
Kong, T., Rosenfeld, A.: Digital topology: introduction and survey. Comput. Gr. Image Process. 48(3), 357–393 (1989)
Lambert, Z., Le Guyader, C., Petitjean, C.: On the inclusion of topological requirements in CNNs for semantic segmentation applied to radiotherapy. In: Calatroni, L., Donatelli, M., Morigi, S., Prato, M, Santacesaria, M. (Eds.) Scale Space and Variational Methods in Computer Vision, pp. 363–375. Springer, Berlin (2023)
Lambert, Z., Petitjean, C., Dubray, B., Ruan, S.: SegTHOR: segmentation of Thoracic Organs at Risk in CT images. In: 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA), pp. 1–6 (2020)
Li, B., Niessen, W.J., Klein, S., Groot, M., Ikram, M.A., Vernooij, M.W., Bron, E.E.: A hybrid deep learning framework for integrated segmentation and registration: evaluation on longitudinal white matter tract changes. In: Medical Image Computing and Computer Assisted Intervention—MICCAI 2019, pp. 645–653 (2019)
Liu, J., Wang, X., Tai, X.C.: Deep convolutional neural networks with spatial regularization, volume and star-shape priori for image segmentation. J. Math. Imaging Vis. 64(6), 625–645 (2022)
Marcus, M., Mizel, V.J.: Transformations by functions in Sobolev spaces and lower semicontinuity for parametric variational problems. Bull. New Ser. Am. Math. Soc. 79(4), 790–795 (1973)
Negrón Marrero, P.: A numerical method for detecting singular minimizers of multidimensional problems in nonlinear elasticity. Numer. Math. 58, 135–144 (1990)
SegTHOR: an ISBI 2019 challenge. https://competitions.codalab.org/competitions/21145
Ségonne, F.: Active contours under topology control-genus preserving level sets. Int. J. Comput. Vis. 79(2), 107–117 (2008)
Ségonne, F., Pacheco, J., Fischl, B.: Geometrically accurate topology-correction of cortical surfaces using nonseparating loops. IEEE Trans. Med. Imaging 26(4), 518–529 (2007)
Shit, S., Paetzold, J.C., Sekuboyina, A., Ezhov, I., Unger, A., Zhylka, A., Pluim, J.P.W., Bauer, U., Menze, B.H.: clDice—a novel topology-preserving loss function for tubular structure segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 16560–16569 (2021)
Vinogradova, K., Dibrov, A., Myers, G.: Towards interpretable semantic segmentation via gradient-weighted class activation map** (student abstract). In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 13943–13944 (2020)
Wirth, B.: On the Gamma-limit of joint image segmentation and registration functionals based on phase fields. Interfaces Free Bound. 18(4), 441–477 (2016)
Xu, Z., Niethammer, M.: DeepAtlas: joint semi-supervised learning of image registration and segmentation. In: Medical Image Computing and Computer Assisted Intervention—MICCAI 2019, pp. 420–429. Springer, Berlin (2019)
Yotter, R.A., Dahnke, R., Thompson, P.M., Gaser, C.: Topological correction of brain surface meshes using spherical harmonics. Hum. Brain Mapp. 32(7), 1109–1124 (2011)
Acknowledgements
This project was co-financed by the French Research National Agency ANR via AAP CE23 MEDISEG ANR project. The authors would like to thank the CRIANN (Centre Régional Informatique et d’Applications Numériques de Normandie, France) for providing computational resources.
Funding
This project was co-financed by the French Research National Agency ANR via AAP CE23 MEDISEG ANR project. The authors would like to thank the CRIANN (Centre Régional Informatique et d’Applications Numériques de Normandie, France) for providing computational resources.
Author information
Authors and Affiliations
Contributions
ZL and CLG have contributed equally to this scientific project and to the writing of this article. ZL prepared the figures. Both authors have reviewed the manuscript.
Corresponding author
Ethics declarations
Conflict of interest
The authors declare they have no financial interests. All authors certify that they have no affiliations with or involvement in any organisation or entity with any financial interest or non-financial interest in the subject matter or materials discussed in this manuscript.
Ethical Approval
The SegTHOR data acquisition was conducted retrospectively using human subject data, made available in open access by the LITIS and the Centre Henri Becquerel (CHB) [28]. The protocol was reviewed and approved by the institution (CHB) board.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Proof of Theorem 1
The proof follows the arguments of the classical direct method of the calculus of variations. Clearly, \(\mathcal {F}(\varphi ) \ge \mu \,\Vert \nabla \varphi \Vert _{L^4(\Omega ,M_2(\mathbb {R}))}^4-4\mu \,{\text{ meas }}(\Omega )\), \(\text{ meas }\) denoting the Lebesgue measure in \(\mathbb {R}^2\). The quantity \(\mathcal {F}(\varphi )\) is thus bounded below by \(-4\mu \,{\text{ meas }}(\Omega )\) and as for \(\varphi ={\text{ Id }}\), \(\mathcal {F}(\varphi )=\frac{\nu }{2}\,\Vert s(\theta )-y\Vert ^2_{L^2(\Omega )}\) is finite (due to the embedding \(BV(\Omega )\subset L^2(\Omega )\), or simply by the fact that \(L^{\infty }(\Omega )\subset L^2(\Omega )\)), the infimum is finite. Let then \((\varphi _k)_k \in \mathcal {W}\) be a minimising sequence, i.e. a sequence such that \(\displaystyle {\lim _{k \rightarrow +\infty }}\,\mathcal {F}(\varphi _k)=\displaystyle {\inf _{\varphi \in \mathcal {W}}}\,\mathcal {F}(\varphi )\). There exists \(K\in \mathbb {N}\) such that \(\forall k \in \mathbb {N}\),
From now on, we assume that \(k \ge K\). According to the coercivity inequality, one gets that \((\varphi _k)_k\) is uniformly bounded with respect to k in \(W^{1,4}(\Omega ,\mathbb {R}^2)\), using the generalised Poincaré inequality [11, pp. 106–107] and the fact that \(\varphi _k={\text{ Id }}\) on \(\partial \Omega \). Besides, \(\det \,\nabla \varphi _k=1\) a.e. on \(\Omega \) so that \(\det \,\nabla \varphi _k\) is uniformly bounded in \(L^{\infty }(\Omega )\). Thus there exists a subsequence—still denoted by \((\varphi _k)_k\)—and \(\bar{\varphi } \in W^{1,4}(\Omega ,\mathbb {R}^2)\) such that
Furthermore, \( \varphi _k \underset{k\rightarrow +\infty }{\rightarrow }\ \bar{\varphi } \) in \(\mathcal {C}^{0,w}(\bar{\Omega },\mathbb {R}^2)\) due to the compact embedding \(W^{1,4}(\Omega ,\mathbb {R}^2) \subset \mathcal {C}^{0,w}(\bar{\Omega },\mathbb {R}^2)\) [8, Sobolev embedding theorem, Theorem 12.12] for any \(w<\frac{1}{2}\). Moreover, there exist a subsequence (common with the previous one, which is always possible)—still denoted by \((\det \,\nabla \varphi _k)_k\)—and \(\delta \in L^{\infty }(\Omega )\) such that
meaning that \(\forall \phi \in L^1(\Omega )\),
As \(\Omega \) is bounded, \(L^2(\Omega ) \subset L^1(\Omega )\) so that \((\det \,\nabla \varphi _k)_k\) weakly converges to \(\delta \) in \(L^2(\Omega )\). By applying [8, Theorem 8.20], we deduce (by uniqueness of the weak limit in \(L^2(\Omega ))\) that \(\det \,\nabla \bar{\varphi }=\delta \) and \(\det \,\nabla \varphi _k \underset{k\rightarrow +\infty }{\overset{*}{\rightharpoonup }} \det \,\nabla \bar{\varphi }\). Then \(\det \,\nabla \bar{\varphi }=1\) a.e.. By continuity of the trace operator [5, Theorem III.9], we get that \(\bar{\varphi }\in {\text{ Id }}+W_0^{1,4}(\Omega ,\mathbb {R}^2)\).
Now, for all \(k\ge K\),
C being a positive constant depending only on \(\Omega \). The assumptions of Ball’s theorems [3, Theorems 1 and 2] thus hold, yielding that \(\varphi _k\) is a homeomorphism of \(\bar{\Omega }\) onto \(\bar{\Omega }\) and \(\varphi _k^{-1}\in W^{1,4}(\Omega ,\mathbb {R}^2)\). The classical change of variable formula holds.
holds for any measurable \(A\subset \bar{\Omega }\) and any measurable function \(f:\mathbb {R}^2 \rightarrow \mathbb {R}\), provided only that one of the integrals exists in the previous relation. The matrix of weak derivatives of \(\varphi _k^{-1}\) is given by \(\nabla \varphi _k^{-1}=\left( \nabla \varphi _k\right) ^{-1} \circ \varphi _k^{-1}\) almost everywhere in \(\Omega \).
The same reasoning applies to \(\bar{\varphi }\) so that it inherits the same smoothness properties: it is a homeomorphism of \(\bar{\Omega }\) onto \(\bar{\Omega }\) and \(\bar{\varphi }^{-1}\in W^{1,4}(\Omega ,\mathbb {R}^2)\).
The stored energy function W is continuous and convex. Let \(\psi _k\) be such that \(\psi _k \underset{k\rightarrow +\infty }{\longrightarrow }\ \bar{\psi }\) in \(W^{1,4}(\Omega ,\mathbb {R}^2)\). Let then \(\psi _{\sigma _1(k)}\) be such that \(I(\psi _{\sigma _1(k)})=\int _{\Omega }\,W(\nabla \psi _{\sigma _1(k)})\,\textrm{d}x \underset{k\rightarrow +\infty }{\longrightarrow }\ \displaystyle {\liminf _{k\rightarrow +\infty } }I(\psi _k)\), which exists by definition of the \(\liminf \). From this subsequence, one can extract a second subsequence \(\psi _{\sigma _1 \circ \sigma _2(k)}\) such that \(\nabla \psi _{\sigma _1 \circ \sigma _2(k)} \underset{k\rightarrow +\infty }{\longrightarrow }\nabla \bar{\psi }\) almost everywhere in \(\Omega \). Applying Fatou’s lemma yields
or rewritten in terms of W,
showing that I is (convex), strongly lower semi-continuous, so, in accordance with [5, Corollary III.8], lower semi-continuous for the weak topology. It remains to study \(\underset{k\rightarrow +\infty }{\lim \inf }\ \,\Vert s(\theta )-y\circ \varphi _k\Vert _{L^2(\Omega )}^2\)—in fact, as will be seen later, we handle \(\underset{k\rightarrow +\infty }{\lim } \,\Vert s(\theta )-y\circ \varphi _k\Vert _{L^2(\Omega )}^2\)—, inspired by prior work by Wirth [33]. We first prove that \(\varphi _k \circ \bar{\varphi }^{-1} \,\underset{k\rightarrow +\infty }{\longrightarrow }\ \,{\text{ Id }}\) in \(\mathcal {C}^{0,\alpha }(\bar{\Omega },\mathbb {R}^2)\) with \(\alpha < \frac{1}{2}\). Recall that (see [8, Definition 12.5]) with \(0<\alpha \le 1, \mathcal {C}^{0,\alpha }(\bar{\Omega },\mathbb {R}^2)\) is the set of functions \(u\in \mathcal {C}^{0}(\bar{\Omega },\mathbb {R}^2)\) such that
It is equipped with the norm
where the notation \(|\cdot |\) is interpreted as the Euclidean norm. Additionally, \(W^{1,4}(\Omega ,\mathbb {R}^2) \subset \mathcal {C}^{0,\lambda } (\bar{\Omega },\mathbb {R}^2)\) [8, Sobolev embedding theorem, Theorem 12.11] for every \(\lambda \in [0,\frac{1}{2}]\) and the embedding is compact for every \(0\le \lambda <\frac{1}{2}\) [8, Rellich-Kondrachov theorem, Theorem 12.12]. Let us take \((\alpha ,\lambda )\) positive real numbers such that \(\left\{ \begin{array}{ccc} \lambda &{}<&{}\frac{1}{2} \\ \frac{\alpha }{\lambda }&{}<&{}\frac{1}{2} \end{array} \right. \), which is always possible. Straightforward computations thus give
From the above, for almost every \(x\in \Omega , \nabla \varphi _k^{-1}=(\nabla \varphi _k)^{-1}(\varphi _k^{-1})\). Additionally, \(\det {\left( \nabla \varphi _k\right) ^{-1}}=1\) almost everywhere. Let \(\mathcal {N}_k\subset \Omega \) be such that \({\text{ meas }}(\mathcal {N}_k)=0\) and for all \(x\in \Omega \backslash \mathcal {N}_k, \det {\left( \nabla \varphi _k\right) ^{-1}}=1\). Let now \(\mathcal {N}_k'\) be such that \(\mathcal {N}_k'=\varphi _k(\mathcal {N}_k)\). From [26, Theorem 1 and Corollary 1], \({\text{ meas }}(\mathcal {N}_k')=0\). For every \(y\notin \mathcal {N}_k', \det {\left( \left( \nabla \varphi _k\right) ^{-1}(\varphi _k^{-1}(y))\right) } =\det {\left( \nabla \varphi _k\right) ^{-1}(x)}\) with \(x \in \Omega \backslash \mathcal {N}_k\), yielding \(\det {\nabla \varphi _k^{-1}}=1\) a.e..
Note first of all that \(y \circ \varphi _k \in L^2(\Omega )\) (the same holds for \(y\circ \bar{\varphi }\) by applying a similar reasoning). Indeed, according to the change of variable formula, valid from Ball’s results (see in particular [3, Proof of Theorem 3]),
the rightmost quantity being uniformly bounded in k, having indeed \(y\in L^{\infty }(\Omega )\) (more precisely, \(y\in BV(\Omega ,\left\{ 1,\ldots ,L\right\} \)) and \(\det {\,\nabla \varphi _k^{-1}}=1\) almost everywhere.
Let now \(\mathcal {O}_i=\left\{ x\in \Omega \,|\,y(x)=i\right\} \) with i any index in \(\left\{ 1,\ldots ,L\right\} \). We first observe that \(\Omega =\displaystyle {\cup _{i=1}^{L}}\,\mathcal {O}_i\cup \mathcal {N}\) with \(\mathcal {N}\) such that \(\text{ meas }(\mathcal {N})=0\).
Let us now consider any measurable set \(\mathcal {S}\subset \Omega \) with Lebesgue measure \(\text{ meas }(\mathcal {S}) \le \eta \). One has, according to the change of variable formula, valid from Ball’s results, and Hölder’s inequality with parameter \(s\ge 2\)
since \(\det {\nabla \bar{\varphi }}=1\) a.e.. Again, due to the smoothness properties of \(\bar{\varphi }\), \(\Omega =\bar{\varphi }^{-1}(\Omega )=\cup _{i=1}^{L}\,\bar{\varphi }^{-1}(\mathcal {O}_i) \cup \bar{\varphi }^{-1}(\mathcal {N})\), which, combined with inequality (9), shows that to achieve our goal, it suffices to prove that \(y \circ \varphi _k \underset{k\rightarrow +\infty }{\longrightarrow }\ y \circ \bar{\varphi }\) in \(L^2(\bar{\varphi }^{-1}\left( \mathcal {O}_i\right) )\). One has:
where \(\Delta \) stands for the symmetric difference. According to the change of variable, valid from Ball’s results, and Hölder’s inequality with parameter \(s\ge 2\),
The rightmost term is uniformly bounded with respect to k since \(\det \,\nabla \varphi _k=1\) a.e.. Besides, as \(\varphi _k \circ \bar{\varphi }^{-1} \,\underset{k\rightarrow +\infty }{\longrightarrow } \,{\text{ Id }}\) in \(\mathcal {C}^{0,\lambda }(\bar{\Omega },\mathbb {R}^2)\), it can be shown (see [33, Footnote 1., p.455]) that \({\text{ meas }}\,\left( \varphi _k \circ \bar{\varphi }^{-1}\,\left( \mathcal {O}_i\right) \Delta \mathcal {O}_i\right) \underset{k\rightarrow +\infty }{\longrightarrow }0\), yielding \(y \circ \varphi _k \underset{k\rightarrow +\infty }{\longrightarrow }\ y \circ \bar{\varphi }\) in \(L^2(\bar{\varphi }^{-1}\left( \mathcal {O}_i\right) )\).
Proof of Theorem 2
The proof is an adaptation of the one of [15, Theorem 2, Chapter 5]. We follow the steps of [15, Theorem 2, Chapter 5] and highlight the main differences [15, Theorem 2, Chapter 5] without dwelling on common elements. In particular, we refer the reader to [15, Theorem 2, Chapter 5] for the proof of (ii).
Let us fix \(\varepsilon >0\). Let m be a positive integer. Let us define the open sets
B(x, r) denoting the open ball with centre x and radius r, and m being chosen large enough so that \(\Vert Dy\Vert (\Omega -\Omega _1)<\varepsilon \).
Set \(\Omega _0 \equiv \emptyset \) and define \(V_k=\Omega _{k+1} {\setminus } \overline{\Omega _{k-1}}, k=1,\ldots \). Let \(\left\{ \zeta _k\right\} _{k=1}^{\infty }\) be a sequence of smooth functions such that
Let \(\rho \in \mathcal {C}_c^{\infty }(\mathbb {R}^2)\) be a mollifier with \({\text{ supp }}\,(\rho )\subset B(0,1), \rho \ge 0\) and \(\int _{\mathbb {R}^2}\,\rho (x)\,\textrm{d}x=1\), and define next \(\rho _{\varepsilon }(x)=\frac{1}{\varepsilon ^2}\,\rho \left( \frac{x}{\varepsilon }\right) \). For each k, select \(\varepsilon _k>0\) so small that
Define \(y_{\varepsilon }=\displaystyle {\sum _{k=1}^{\infty }}\,\rho _{\varepsilon _k}*\left( y\zeta _k\right) \). By construction of the sets \(\left\{ \Omega _k\right\} _k\), there exists \(n_0\) such that \(\text{ ess }\,\text{ supp }\,(y)\subset \Omega '\subset \subset \Omega _{n_0}\). The quantity \(\displaystyle {\sum _{k=1}^{\infty }}\,\rho _{\varepsilon _k}*\left( y\zeta _k\right) \) can thus be restricted to the indices \(\displaystyle {\sum _{k=1}^{n_0}}\,\rho _{\varepsilon _k}*\left( y\zeta _k\right) \), yielding \(y_{\varepsilon }=\displaystyle {\sum _{k=1}^{n_0}}\,\rho _{\varepsilon _k}*\left( y\zeta _k\right) \in \mathcal {C}^{\infty }(\Omega )\) with \({\text{ supp }}(y_{\varepsilon }) \subset \Omega _{n_0+1}\), set that depends on \(\varepsilon \). Note that taking \(\tilde{\varepsilon }\le \varepsilon \) would generate another partitioning \(\tilde{\Omega }_{k}\) that may be obtained by relabelling the partitioning \({\Omega }_{k}\).
Since \(y=\displaystyle {\sum _{k=1}^{n_0}}y\zeta _k\), using Minkowski inequality,
Consequently, \(y_{\varepsilon } \, \underset{}{\longrightarrow }\,y\) in \(L^p(\Omega )\) as \(\varepsilon \) tends to 0 (and in particular, in \(L^1(\Omega )\)). Thus, as \(\Vert Dy\Vert (\Omega ) \le \displaystyle {\liminf _{\varepsilon \rightarrow 0}}\,\Vert Dy_{\varepsilon }(\Omega )\Vert \), it suffices to prove that \(\displaystyle {\limsup _{\varepsilon \rightarrow 0}}\,\Vert Dy_{\varepsilon }(\Omega )\Vert \le \Vert Dy\Vert (\Omega )\), which is led to the reader following [15, Theorem 2, Chapter 5].
Now, if \(x\in \Omega _1 \subset \subset \Omega _2\) then
yielding \(|y_{\varepsilon }(x)| \le (L-1)\,\int _{}\rho _{\varepsilon _1}(x-s)\,\textrm{d}s=L-1\).
If \(x \in \Omega _l{\setminus } \overline{\Omega _{l-1}}\) with \(l>1\), then
showing that \(y_{\varepsilon }\) is uniformly bounded in the \(L^{\infty }\)-norm (the bound depending only on L which is fixed).
Proof of Theorem 3
First observe that with \(\varphi _l\) defined above,
\(E_{l,j}(\varphi ,V,W)\ge 0\) and taking \(\varphi ={\text{ Id }}\), \(V=I_2\) and \(W=I_2\) shows that the functional is proper since \(s(\theta )\in L^2(\Omega )\) and \(y_l\in \mathcal {C}^{\infty }(\Omega )\) (bounded uniformly with respect to l), and subsequently, that the infimum is finite.
Let us denote by \(\left( \varphi _{l,j,k},V_{l,j,k},W_{l,j,k}\right) \) a minimising sequence of the decoupled problem \((DP)_{l,j}\), that is, a sequence such that
In particular,
Let us take in particular \(\varepsilon =\frac{1}{\gamma _j}\). Then there exists \(N_{l,j}\in \mathbb {N}\) such that \(\forall k \in \mathbb {N}\),
We now take \(k:=N_{l,j}\) and for the sake of readability, we denote by \(\varphi _{l,j}:=\varphi _{l,j,N_{l,j}}\) and similarly for \(V_{l,j}\) and \(W_{l,j}\). Using the inequality \((a-b)^2\ge \frac{1}{2}a^2-b^2\), a first coercivity-type inequality can be obtained, of the form
showing that
Thus there exist a common subsequence, obtained by applying a diagonal extraction procedure, denoted by \((V_{l,\Psi (j)})\) (respectively, \((\det {V_{l,\Psi (j)}})\) and \(\bar{V}_l \in L^4(\Omega ,M_2(\mathbb {R}))\) and \(\bar{\delta }_l\in L^2(\Omega )\) such that
In addition,
yielding
The sequence \((\varphi _{l,\Psi (j)})\) is thus uniformly bounded in \(W^{1,2}(\Omega ,\mathbb {R}^2)\) according to the generalised Poincaré inequality, and there exist a subsequence obtained by applying a diagonal extraction procedure and still denoted by \((\varphi _{l,\Psi (j)})\) and \(\bar{\varphi }_l\in W^{1,2}(\Omega ,\mathbb {R}^2)\) such that
The same kind of argument applies to \((W_{l,\Psi (j)})\) yielding
Additionally, \((\det {W_{l,\Psi (j)}})\) is uniformly bounded in \(L^{\infty }\)-norm and so there exists \(\bar{\bar{\delta }}_l\in L^{\infty }(\Omega )\) such that
and finally, \(\bar{\bar{\delta }}_l=1\) a.e..
Let us now set \(x_{l,\Psi (j)}=V_{l,\Psi (j)}-\nabla \varphi _{l,\Psi (j)}\).
On the one hand, since \(\Vert V_{l,\Psi (j)}-\nabla \varphi _{l,\Psi (j)}\Vert _{L^2(\Omega ,M_2(\mathbb {R}))}\le \frac{2}{\gamma _{\Psi (j)}}\,\left[ \inf _{\varphi \in \mathcal {W}}\,\mathcal {F}(\varphi )+\frac{C+1}{l}+\frac{1}{\gamma _0}\right] \), it implies that \(x_{l,\Psi (j)} \underset{j \rightarrow +\infty }{\longrightarrow }\ 0 \) in \(L^2(\Omega ,M_2(\mathbb {R}))\). Consequently, \(V_{l,\Psi (j)}-\nabla \varphi _{l,\Psi (j)} \underset{j \rightarrow +\infty }{\rightharpoonup }\ 0\) in \(L^2(\Omega ,M_2(\mathbb {R}))\), yielding \(\nabla \varphi _{l,\Psi (j)} \underset{j \rightarrow +\infty }{\rightharpoonup }\ \bar{V}_l\) in \(L^2(\Omega ,M_2(\mathbb {R}))\). Indeed, \(\forall \phi \in L^2(\Omega ,M_2(\mathbb {R}))\), \(\int _{\Omega }\,x_{l,\Psi (j)}\,:\,\phi \,\textrm{d}x \underset{j \rightarrow +\infty }{\rightarrow }\ 0\), that is, \(\int _{\Omega }\,V_{l,\Psi (j)}-\nabla \varphi _{l,\Psi (j)}\,:\,\phi \,\textrm{d}x \underset{j \rightarrow +\infty }{\rightarrow }\ 0\). But \(V_{l,\Psi (j)}\underset{j \rightarrow +\infty }{\rightharpoonup }\ \bar{V}_l\) in \(L^4(\Omega ,M_2(\mathbb {R}))\) so in \(L^2(\Omega ,M_2(\mathbb {R}))\), which shows that, \(\forall \phi \in L^2(\Omega ,M_2(\mathbb {R}))\),
Then \(\nabla \varphi _{l,\Psi (j)} \underset{j \rightarrow +\infty }{\rightharpoonup }\ \bar{V}_l\) in \(L^2(\Omega ,M_2(\mathbb {R}))\). On the other hand, \(\nabla \varphi _{l,\Psi (j)} \underset{j \rightarrow +\infty }{\rightharpoonup }\ \nabla \bar{\varphi }_l\) in \(L^2(\Omega ,M_2(\mathbb {R}))\), which, by uniqueness of the weak limit in \(L^2(\Omega ,M_2(\mathbb {R}))\), enables one to conclude that \(\nabla \bar{\varphi }_l=\bar{V}_l\in L^4(\Omega ,M_2(\mathbb {R}))\).
Let us now focus on \((\det {V_{l,\Psi (j)}})\). One has \(\det {V_{l,\Psi (j)}}=\det {\nabla \varphi _{l,\Psi (j)}}+d_{l,\Psi (j)}\) with
\(\left( x_{l,\Psi (j)}\right) _{k,n}\) being the element of the \(k^{\textrm{th}}\) row and \(n^{\textrm{th}}\) column of the matrix \(x_{l,\Psi (j)}\) and with \(\varphi _{l,\Psi (j)}=\left( \varphi _{l,\Psi (j)}^1,\varphi _{l,\Psi (j)}^2\right) \). Then it holds that
The quantity \(\Vert \nabla \varphi _{l,\Psi (j)}\Vert _{L^2(\Omega ,M_2(\mathbb {R}))}\) being bounded independently of j and l and having \(x_{l,\Psi (j)}\underset{j \rightarrow +\infty }{\longrightarrow }\ 0 \) in \(L^2(\Omega ,M_2(\mathbb {R}))\), it follows that \( d_{l,\Psi (j)}\underset{j \rightarrow +\infty }{\longrightarrow }\ 0 \) in \(L^1(\Omega )\).
From [8, Theorem 1.14], if \(\Psi _j \,\,\underset{j \rightarrow +\infty }{\rightharpoonup }\ \bar{\Psi }\) in \(W^{1,2}(\Omega ,\mathbb {R}^2)\), then \(\det {\nabla \Psi _j}\,\,\underset{j \rightarrow +\infty }{\longrightarrow } \det {\nabla \bar{\Psi }} \) in the sense of distributions. On the one hand, \(\forall \phi \in \mathcal {D}(\Omega )\),
as \((\det {V_{l,\Psi (j)}})\) weakly converges to \(\overline{\delta }_l\) in \(L^2(\Omega )\) as j tends to \(+\infty \). On the other hand,
with \({\int _{\Omega }}\,\det {\nabla \varphi _{l,\Psi (j)}}\,\phi \,\textrm{d}x \,\,\underset{j \rightarrow +\infty }{\longrightarrow } \,\,\int _{\Omega }\,\det {\nabla \bar{\varphi }_l}\,\phi \,\textrm{d}x\), as \(\det {\nabla \varphi _{l,\Psi (j)}}\) converges to \(\det {\nabla \bar{\varphi }_l}\) in the sense of distributions when j tends to \(+\infty \) and \(\left| \int _{\Omega }\,d_{l,\Psi (j)}\,\phi \,\textrm{d}x\right| \le \Vert d_{l,\Psi (j)}\Vert _{L^1(\Omega )}\Vert \phi \Vert _{\mathcal {C}^0(\bar{\Omega })}\) \(\underset{j\rightarrow +\infty }{\rightarrow }0\). Consequently, \(L^2(\Omega ) \ni \det {\nabla \bar{\varphi }_l}=\bar{\delta }_l\in L^2(\Omega )\) in the sense of distributions and finally \(\det {\nabla \bar{\varphi }_l}=\bar{\delta }_l\) a.e..
A similar reasoning applies to handle the component \(\Vert W_{l,\Psi (j)}-\nabla \varphi _{l,\Psi (j)}\Vert ^2_{L^2(\Omega ,M_2(\mathbb {R}))}\) and leads to
Invoking again the generalised Poincaré inequality, it comes that \(\bar{\varphi }_l\in {\text{ Id }}+W_0^{1,4}(\Omega ,\mathbb {R}^2)\) with \(\det {\nabla \bar{\varphi }_l}=1\) a.e.. Subsequently, still following Ball’s results, \(\bar{\varphi }_l\) is a homeomorphism of \(\bar{\Omega }\) to \(\bar{\Omega }\) and \(\bar{\varphi }_l^{-1}\in W^{1,4}(\Omega ,\mathbb {R}^2)\).
Using the Lipschitz property of \(Ey_l\) and Rellich-Kondrachov theorem ([5, Theorem IX.16]) that gives the compact embedding \(W^{1,2}(\Omega ,\mathbb {R}^2) \subset L^2(\Omega )\), it can be observed that \(\Vert s(\theta )-(Ey_l)\circ \varphi _{l,\Psi (j)}\Vert _{L^2(\Omega )}^2\,\,\underset{j \rightarrow +\infty }{\longrightarrow } \,\,\Vert s(\theta )-(Ey_l)\circ \bar{\varphi }_{l}\Vert _{L^2(\Omega )}^2\), so that going back to (10), it yields
Thus
and finally, as \(\det {\nabla \bar{\varphi }_l}=1\) a.e.,
Observe that \(\forall \varphi \in {\text{ Id }}+W_{0}^{1,4}(\Omega ,\mathbb {R}^2)\) with \(\det {\nabla \varphi }=1\) a.e.,
which enables one to conclude that \(\bar{\varphi }_l\) is a minimiser of \(\mu \,\Vert \nabla \cdot \Vert _{L^4(\Omega ,M_2(\mathbb {R}))}^4+\dfrac{\nu }{2}\,\Vert s(\theta )-y_l\circ \cdot \Vert _{L^2(\Omega )}^2\), after passing to the \(\liminf \) in the previous inequality when j tends to \(+\infty \). Reasoning as before,
so that
with \(C:=\frac{C\nu }{2}\) and
Passing to the limit when l tends to \(+\infty \) yields
and the result follows, after noticing that
yielding
About this article
Cite this article
Lambert, Z., Le Guyader, C. About the Incorporation of Topological Prescriptions in CNNs for Medical Image Semantic Segmentation. J Math Imaging Vis (2024). https://doi.org/10.1007/s10851-024-01172-3
Received:
Accepted:
Published:
DOI: https://doi.org/10.1007/s10851-024-01172-3