3.1 Research program

The recent advent of an increasing number of new microscopy techniques giving access to high throughput screenings and micro or nano-metric resolutions provides a means for quantitative imaging of biological structures and phenomena. To conduct quantitative biological studies based on these new data, it is necessary to develop non-standard specific tools. This requires using a multi-disciplinary approach. We need biologists to define experiment protocols and interpret the results, but also physicists to model the sensors, computer scientists to develop algorithms and mathematicians to model the resulting information. These different expertises are combined within the Morpheme team. This generates a fecund frame for exchanging expertise, knowledge, leading to an optimal framework for the different tasks (imaging, image analysis, classification, modeling). We thus aim at providing adapted and robust tools required to describe, explain and model fundamental phenomena underlying the morphogenesis of cellular and supra-cellular biological structures. Combining experimental manipulations, in vivo imaging, image processing and computational modeling, we plan to provide methods for the quantitative analysis of the morphological changes that occur during development. This is of key importance as the morphology and topology of mesoscopic structures govern organ and cell function. Alterations in the genetic programs underlying cellular morphogenesis have been linked to a range of pathologies.

Biological questions we will focus on include:

1. what are the parameters and the factors controlling the establishment of ramified structures? (Are they really organized to ensure maximal coverage? How are genetic and physical constraints limiting their morphology?),
2. how are newly generated cells incorporated into reorganizing tissues during development? (is the relative position of cells governed by the lineage they belong to?)

Our goal is to characterize different populations or development conditions based on the shape of cellular and supra-cellular structures, e.g. micro-vascular networks, dendrite/axon networks, tissues from 2D, 2D+t, 3D or 3D+t images (obtained with confocal microscopy, video-microscopy, photon-microscopy or micro-tomography). We plan to extract shapes or quantitative parameters to characterize the morphometric properties of different samples. On the one hand, we will propose numerical and biological models explaining the temporal evolution of the sample, and on the other hand, we will statistically analyze shapes and complex structures to identify relevant markers for classification purposes. This should contribute to a better understanding of the development of normal tissues but also to a characterization at the supra-cellular scale of different pathologies such as Alzheimer, cancer, diabetes, or the Fragile X Syndrome. In this multidisciplinary context, several challenges have to be faced. The expertise of biologists concerning sample generation, as well as optimization of experimental protocols and imaging conditions, is of course crucial. However, the imaging protocols optimized for a qualitative analysis may be sub-optimal for quantitative biology. Second, sample imaging is only a first step, as we need to extract quantitative information. Achieving quantitative imaging remains an open issue in biology, and requires close interactions between biologists, computer scientists and applied mathematicians. On the one hand, experimental and imaging protocols should integrate constraints from the downstream computer-assisted analysis, yielding to a trade-off between qualitative optimized and quantitative optimized protocols. On the other hand, computer analysis should integrate constraints specific to the biological problem, from acquisition to quantitative information extraction. There is therefore a need of specificity for embedding precise biological information for a given task. Besides, a level of generality is also desirable for addressing data from different teams acquired with different protocols and/or sensors. The mathematical modeling of the physics of the acquisition system will yield higher performance reconstruction/restoration algorithms in terms of accuracy. Therefore, physicists and computer scientists have to work together. Quantitative information extraction also has to deal with both the complexity of the structures of interest (e.g., very dense network, small structure detection in a volume, multiscale behavior, $...$) and the unavoidable defects of in vivo imaging (artifacts, missing data, $...$). Incorporating biological expertise in model-based segmentation methods provides the required specificity while robustness gained from a methodological analysis increases the generality. Finally, beyond image processing, we aim at quantifying and then statistically analyzing shapes and complex structures (e.g., neuronal or vascular networks), static or in evolution, taking into account variability. In this context, learning methods will be developed for determining (dis)similarity measures between two samples or for determining directly a classification rule using discriminative models, generative models, or hybrid models. Besides, some metrics for comparing, classifying and characterizing objects under study are necessary. We will construct such metrics for biological structures such as neuronal or vascular networks. Attention will be paid to computational cost and scalability of the developed algorithms: biological experimentations generally yield huge data sets resulting from high throughput screenings. The research of Morpheme will be developed along the following axes:

• Imaging: this includes i) definition of the studied populations (experimental conditions) and preparation of samples, ii) definition of relevant quantitative characteristics and optimized acquisition protocol (staining, imaging, $...$) for the specific biological question, and iii) reconstruction/restoration of native data to improve the image readability and interpretation.
• Feature extraction: this consists in detecting and delineating the biological structures of interest from images. Embedding biological properties in the algorithms and models is a key issue. Two main challenges are the variability, both in shape and scale, of biological structures and the huge size of data sets. Following features along time will allow to address morphogenesis and structure development.
• Classification/Interpretation: considering a database of images containing different populations, we can infer the parameters associated with a given model on each dataset from which the biological structure under study has been extracted. We plan to define classification schemes for characterizing the different populations based either on the model parameters, or on some specific metric between the extracted structures.
• Modeling: two aspects will be considered. This first one consists in modeling biological phenomena such as axon growing or network topology in different contexts. One main advantage of our team is the possibility to use the image information for calibrating and/or validating the biological models. Calibration induces parameter inference as a main challenge. The second aspect consists in using a prior based on biological properties for extracting relevant information from images. Here again, combining biology and computer science expertise is a key point.

5.1 Video

A presentation video has been published on the CNRS Youtube channel concerning the work of Laure Blanc-Féraud.

6.1 New software

7.1 Fluorescence image deconvolution microscopy via generative adversarial learning (FluoGAN)

Participants: Hamza Mentagui, Laure Blanc-Féraud, Luca Calatroni, Sébastien Schaub.

Image super-resolution techniques exploiting the stochastic fluctuations of image intensities have become a powerful tool in fluorescence microscopy. Compared to other approaches, these techniques can be applied under standard acquisition settings and do not require special microscopes nor fluorophores. Most of these approaches can be mathematically modeled making use of second-order statistics possibly combined with a priori regularization on the desired solution. In 15, we consider a different paradigm and formulate a physical-inspired data-driven approach based on generative learning. By simulating fluorescence and noise fluctuations by means of a suitable double Poisson-type process, the unknown distribution of the fluctuating sequence of low- resolution and noisy images is approximated via a GAN-type approach (Generative Adversial Network) where both physical and network parameters are optimized, see 2. We also provide theoretical insights on the choice of the corresponding cost functionals and gradient computations, and assess practical performance on simulated Argolight (see Figure 1).

(a) Temporal mean $\overline{𝐲}$ of fluctuating stack simulated by SOFI tool.

(b) Ground Truth, $d=120nm$.

 

(c) Simulated background.

 

(c) FluoGAN result.

7.2 From Photon Emission to Super-Resolution: The MUFASA Simulator

Participants: Wessim Omezzine, Hamza Mentagui, Laure Blanc-Féraud, Luca Calatroni, Sébastien Schaub.

Advances in Super-resolution techniques, like Single Molecule Localization Microscopy (SMLM) and Fluorescence Fluctuation-based SRM (FF-SRM) allowed to go beyond the diffraction limit. These breakthroughts rely on the photophysical processes, including the stochastic nature of photon emission and transitions between quantum states. Despite this progress, replicating these protocols in experimental setups remains challenging due to the complex behavior of fluorophores, which depends on environmental factors like laser power and molecular properties. Existing simulators often offer only one type of protocol and focus primarily on ”on” and ”off” regimes, failing to capture the fluctuation characteristics of fluorescence. We introduce MUFASA (Muitl-protocol Unified Fluorescence-based Advanced Simula), a comprehensive simulator that unifies multiple protocols within a physically grounded framework. Using Continuous Time Markov Chains, MUFASA captures detailed quantum state transitions and stochastic photon emission dynamics. Unlike existing simulators, MUFASA incorporates environment-dependent transition rates and continuous emission dynamics, providing enhanced realism and flexibility. MUFASA supports key protocols, including STORM, PALM, Blinking and FFSRM, offering insights into molecular dynamics and optimizing experimental conditions. Using Wasserstein distances, we validate the distribution of MUFASA simulations against experimental data, demonstrating superior accuracy compared to existing tools. Its graphical interface enables accessible customization, empowering researchers to simulate both single-molecule behavior and complex biological structures. By bridging theoretical modeling with practical experimentation, MUFASA advances fluorescence microscopy simulation, facilitating the development and optimization of imaging techniques and reconstruction algorithms. Moreover, it serves as a valuable tool for generating realistic data to train and evaluate learning models, further enhancing research capabilities.

7.3 Off-the-grid dynamic super-resolution for fluorescence microscopy

Participants: Aneva Tsafack, Laure Blanc-Féraud, Gilles Aubert.

Recently, the Curves Represented On Charges (CROC) functional

was introduced for curve reconstruction in microscopy images, optimizing within the space of finite vector-valued Radon measures with finite divergence, denoted $𝒱\left(𝒳\right)$. A key question was how to effectively compute solutions of this functional. To address this, a Sliding Frank-Wolfe algorithm adapted to curves was implemented, but it involved a challenging non-convex optimization step in the space of 1-rectifiable curves. In this work, we propose a relaxation of the functional, enabling the use of classical convex optimization algorithms for implementation.

We define a space of Lebesgue measures

and prove its density in the previous space $𝒱\left(𝒳\right)$ under the weak-topology. Reformulating the CROC functional in this space, we prove the convergence of minimizing sequences to solutions of the original functional.

For numerical implementation, a key challenge was modeling the data term due to the mismatch between the vector-valued optimization framework and the scalar nature of the observed images. Initial attempts to derive observed vector fields from image gradients proved unsatisfactory. To address this, a scalar-based acquisition model was developed to generate scalar images from velocity vector fields along curves:

where $\left|m\right|$ is the total variation measure of $m$ and $h$ the blur kernel. While this new functional is non-convex, its relaxation in $𝒲\left(𝒳\right)$ can be discretized and efficiently optimized using standard algorithms, yielding promising numerical results, as illustrated in Figure 2.

7.4 Off-the-grid super-resolution for Poisson data

Participants: Luca Calatroni.

This work was made in collaboration with M. Lazzaretti and C. Estatico (University of Genova, Italy).

To model acquisitions where the observed noise is due to low photon-counting processes (which is often the case in microscopy applications), we consider a variational model where off-the-grid Total Variation regularization is coupled with a Kullback-Leibler data term under a non-negativity constraint. Analytically, we study the optimality conditions of the composite functional and analyze its dual problem. Then, we consider an homotopy strategy to select an optimal regularization parameter and use it within a Sliding Frank-Wolfe algorithm. Several numerical experiments on both 1D/2D/3D simulated and real 3D fluorescent microscopy data are reported on Figure 3

This work is reported in the preprint 21 and part of the PhD thesis 18.

7.5 Patch-based learning of adaptive parameter maps for image reconstruction

Participants: Luca Calatroni, Xavier Descombes, Claudio Fantasia.

In collaboration with Rim Rekik Dit Nekhili (Inria).

In this project, we consider a variational reconstruction model based on a Total Variation regularization combined with a mixed data-fidelity term adapted to the presence of either Gaussian or Poisson noise in the observed data. To adjust the effect of TV (Total Variation)regularization against data fidelity, a non-stationary (i.e., pixel-adaptive) parameter map is considered and learned patch-wise by means of a light neural network taking as input a noisy patch and providing as output a locally optimal parameter. Training is thus performed by providing as labels pairs $(y_i,{\widehat{\lambda }}_i)$ where $y_i$ is an image patch and $\widehat{{\lambda }_i}$ is an optimal parameter found by golden-section and optimized over the SSIM score. The computed parameter maps at inference time provide visual description of local features (textures VS. homogeneous areas) and the reconstructions computed are therefore more reliable than the ones obtained by scalar regularization. To select which model is to be used during processing (${\ell }_2$ square fidelity for Gaussian noise or Kullback-Leibler divergence for Poisson noise), a binary classification network is trained still on patches. Overall, the proposed procedure is therefore blind to the noise distribution and adaptive to local features and combines deep learning and image processing in an interpretable way (see some results on Figure 4).

   

7.6 Deep-equilibrium approaches for learning regularization in Poisson inverse problems

Participants: Luca Calatroni, Christian Daniele.

This work was made in collaboration with S. Villa (University of Genoa) and S. Vaiter (LJAD, CNRS).

We consider the framework of deep equilibrium models to learn image regularizers in the context of imaging data corrupted by Poisson noise. In this framework, the Kullback-Leibler divergence is usually considered as a data term, which poses some difficulties due to its lack of Lipschitz-smoothness around zero. By using a mirror descent algorithm and enforcing a Lipschitz-like condition to guarantee convergence of the scheme even in non $L$-smooth settings, we propose a training strategy that learns efficiently the regularizer using limited data and reduced computational times. The key ingredients of the proposed approach are an efficient forward step obtained by means of an efficient backtracking strategy and a cheap backward step relying on Jacobian-free approximations. Some encouraging preliminary results on image denoising and image deblurring show that the proposed approach improves significantly upon well-known approaches (for instance, expectation-maximization, TV regularization...) classically used to process Poisson data (see Figure 5)

7.7 ${\ell }_0$ continuous Bregman-relaxations for sparse problems with non quadratic data terms

Participants: Luca Calatroni, M'hamed Essafri.

This work was made in collaboration with E. Soubies (IRIT, CNRS).

We consider the minimization of ${\ell }_0$-regularized criteria involving non-quadratic data terms such as the Kullback-Leibler divergence and the logistic regression, possibly combined with an ${\ell }_2$ regularization. We first prove the existence of global minimizers for such problems and characterize their local minimizers. Then, we propose a new class of continuous relaxations of the ${\ell }_0$ pseudo-norm, termed as ${\ell }_0$ Bregman Relaxations (B-rex). They are defined in terms of suitable Bregman distances and lead to exact continuous relaxations of the original ${\ell }_0$-regularized problem in the sense that they do not alter its set of global minimizers and reduce its non-convexity by eliminating certain local minimizers. Both features make such relaxed problems more amenable to be solved by standard non-convex optimization algorithms. In this spirit, we consider the proximal gradient algorithm and provide explicit computation of proximal points for the B-rex penalty in several cases. Finally, we report a set of numerical results illustrating the geometrical behavior of the proposed B-rex penalty for different choices of the underlying Bregman distance, its relation with convex envelopes, as well as its exact relaxation properties in 1D/2D and higher dimensions. For the particular case of the KL divergence, tigher relaxations can be found as well as tailored Bregman-proximal gradient algorithms minimizing the relaxed critera efficiently (see Figure 6).

A preprint of this work is online at 20 and the study on the KL case is published in 12.

7.8 Parameter Free FISTA

Participants: Luca Calatroni.

This work was made in collaboration with A. Rondepierre, C. Dossal (IMT et INSA, Toulouse), J.F. Aujol (IMB, Bordeaux) and H. Labarriere (University of Genoa).

We consider a parameter-free version of the famous FISTA algorithm to solve composite non-smooth optimization problems ubiquitous in the field of regularized inverse problems and machine learning. The technique relies on the use of a combined restarting and adaptive backtracking strategy. Several variants of FISTA enjoy a provable linear convergence rate for the function values of the form under the prior knowledge of problem conditioning. These parameters are nonetheless hard to estimate in many practical cases. Recent works address the problem by estimating parameters in two steps via suitable adaptive strategies. In our work, both parameters can be estimated at the same time by means of an algorithmic restarting scheme where, at each restart, a nonmonotone estimation is performed. For this scheme, theoretical convergence results are proved, showing that linear convergence can be achieved along with quantitative estimates of the conditioning. The resulting free-FISTA is therefore parameter-free. Several numerical results are reported to confirm the practical interest of its use in many example problems (see Figure 7).

This project has been published in 7.

7.9 Dynamic analysis framework for live-cell imaging to detect cell responses to anticancer drugs using image and signal processing, and deep learning

Participants: Asma Chalabi, Eric Debreuve.

This work was conducted in collaboration with Jérémie Roux, IPMC, Sophia Antipolis, France.

Cell division and cell death are the main indicators to evaluate cancer drug action, and only their accurate measures can reveal the actual potency and efficacy of a compound. The detection of cell division and cell death events in live-cell assays has the potential to produce robust metrics of drug pharmacodynamics and return a more comprehensive understanding of tumor cells responses to cancer therapeutic combinations. Knowing precisely when a cell death or a cell division occurs in a live-cell experiment allows to study the relative contribution of different drug effects, such as cytotoxic or cytostatic effects, on a cell population. Yet, classical methods require dyes to measure cell viability as an end-point assay with whole population counts, where the proliferation rates can only be estimated when both viable and dead cells are labeled simultaneously.

Live-cell imaging is a promising cell-based assay to determine drug efficacies, with the main limitation being the accuracy and depth of the analyses to detect and predict automatically cellular response phenotypes (cell death and division, which share some morphological features). This work proposes a method integrating deep learning using neural networks, and image and signal processing to perform dynamic image analyses of single-cell events in time-lapse microscopy experiments of drug pharmacological profiling. The method works by automatically tracking the cells, extracting radiometric and morphologic cell features, and analyzing the temporal evolution of these features for each cell to detect cellular events such as division and cell death, as well as acquiring signaling pathway dynamics (see Figure 8). A case of study comprising the analyses of caspase-8 single-cell dynamics and other cell responses to cancer drugs has been achieved. The aim was to perform automatically, at a large scale, the necessary analyses to augment a previous phenotype prediction method and to apply it to various cancer cell lines of a human cancer cell line panel to improve our live-cell OMICS profiling approaches, and, in a longer term, to scale up pharmacological screening of new cancer drugs.

7.10 Identifying autofluorescence in biological samples using hyperspectral imaging

Participants: Eric Debreuve.

This work was made in collaboration with Sébastien Schaub and Jenifer Croce, IMEV, Villefranche-sur-mer, France.

Fluorescence imaging of marine samples (animals or plants) remains a challenge due to the inevitable endogenous fluorescence (or autofluorescence) common in these samples, for example due to an animal ingesting algae exhibiting endogenous fluorescence. The autofluoresence superimposes with the fluorescence of the probes which are the target of a specific study. The aim of this work is to take advantage of recent improvements in fluorescence imaging to identify and subtract sample autofluorescence from probe fluorescence using hyperspectral imaging, i.e. using so-called $\Lambda$$\lambda$ acquisitions (confocal acquisitions varying both excitation and detection wavelengths). This requires to identify the various excitation and emission spectra, and the corresponding concentration maps. A first approach using blind source separation by ICA (Independent Component Analysis) has been developed. Initial results are encouraging (see Figure 9). The challenge now is to improve the approach by imposing constraints linked to the physics of the problem, notably the fact that the emission wavelength is necessarily larger than the excitation wavelength.

7.11 Machine learning for organoid phenotyping

Participants: Alexandre Martin, Xavier Descombes, Ivan Magistro Contenta, Stephan Clavel [IPMC, Sophia Antipolis], Xavier Coumoul [Metatox, Paris].

The aim of this work is to built a phenotypic map of organoids. We model an organoid as a 3D shape containing several layers of cells composed of 2D point patterns on the sphere. Our goal is to classify organoids with respect to these three components.

The main work during this year concerns machine learning for graphs in order to combine discriminative capacity and interpretability. We advanced on organoid classification by integrating more complex neural network architectures. We first considered Convolutional Neural Networks (CNNs) to extract spatial features from organoid images. Although effective, CNNs struggled to capture the relational and structural complexity of cellular arrangements. We therefore moved to Graph Neural Networks (GNNs), representing the data as graphs with nodes for cells and edges for spatial or functional relationships. Early experiments with Graph Convolutional Networks (GCNs) showed promise but lacked the capacity to handle the geometric properties of the data. To address this bottleneck, we employed geometric GNNs, including Equivariant Neural Networks (ENNs), which respect rotational and translational symmetries. These models significantly improved the classification accuracy by taking advantage of the invariances inherent to biological structures. This progression in architecture has improved both the performance and biological relevance of the results. Furthermore, data augmentation techniques were applied to enhance model robustness and performance.

The organoid modeling is based on nuclei detection, each nuclei representing a graph node. In previous year, we have compared different approaches to detect nuclei. Cellpose, a state of the art deep learning model, has provided the best performances in terms of accuracy on 2D slices (see Figure 10). The performances on 3D datasets are similar but the computation time is heavy (about 20 min for one volume). Therefore, we have studied different frugal approaches to reduce the model size, including pruned and dilated models. Currently we investigate the distillation process by considering a teacher-student framework. The first results are encouraging.

7.12 Detection and analysis of pores in the basal lamina of breast cancer spheroids

Participants: Zhenyu Zhu, Xavier Descombes, Frédéric Luton [IPMC, Sophia Antipolis].

The pores are detected on a 2D image of Collagen IV. By comparing on six annoted samples a geometric method based on marked point process and the neuronal networks, Cellpose and Stardist, the most efficient method and the appropriate settings have been selected. After fine-tuning the model, Cellpose appears to be better suited to detecting pores (see Figure 11). Based on the detection results, we can establish statistics on the size and number of pores. In a next step, we considered the structure of Collagen in 3D. The pore detection is still performed in 2D and the pores in 3D are obtained by aggregating the 2D slices. To recover the local geometry in order to have a precise estimation of the pore size, we compute an alpha shape of the spheroid and project the detected 3D pores on the local tangent plane.

A work in progress consists then to detect filipodes representing cell extensions through the pores.

7.13 An automatic BANFF classification

Participants: Meryem Sikouky, Mohammad Mohammad, Xavier Descombes, Damien Ambrosetti [CHU, Nice], Giorgio Toni [CHU, Nice], Paul Hannetel [CHU, Nice].

The Banff classification is a standardized system used to assess transplant rejection risks from kidney biopsies. It involves the examination of whole-slide images (WSIs) by pathologists to identify and score various histological features. Automating this process through image analysis and machine learning could potentially improve the efficiency and objectivity of the evaluation. As a first step towards this goal, this project focuses on the segmentation of kidney histology images. The aim is to accurately identify and delineate different anatomical structures within the kidney, such as glomeruli, PTC, vessels, and tubules. The initial phase of the research involved exploring various neural network architectures for this segmentation task, with the Attention Unet model proving to be the most effective as shown for the example of glomeruli segmentation in Figure 12 (top panel). Building upon this, the project is now investigating a multi-architecture approach for segmentation. Instead of using a single model to segment all kidney structures, the idea is to employ specialized architectures, each dedicated to segmenting a specific structure. This approach aims to leverage the strengths of individual architectures and enhance the overall accuracy of the segmentation process. The next challenge lies in effectively combining the outputs from these individual architectures to generate a comprehensive segmentation map encompassing all kidney structures. To achieve this, the project is exploring various fusion mechanisms, including both neural network-based architectures (mainly the self-supervised learning paradigm) and traditional statistical methods. While the initial work focused on segmenting image patches, the ultimate goal is to extend this capability to whole-slide images (WSIs) (example in Figure 12, bottom panel). WSIs are the standard format used by pathologists for diagnosis, and the ability to segment these images directly will significantly enhance the clinical applicability of the research.

Concerning the work on glomeruli detection conducted last year, we have executed the experiments on a Paris cohort to test the multicentric property of our approach. The accuracy is slighlty reduced compared to that obtained on the Nice cohort used for the training. However, we were able to increase this accuracy after a domain transfer obtained with a generative adversial network (GAN).

7.14 Reinforcement Learning in histopathology

Participants: Mohamad Mohamad, Xavier Descombes, Damien Ambrosetti [CHU, Nice], Nicolas Pote [Hôpital Bichat, Paris], Maxime Gassier [Hôpital Bichat, Paris].

Drawing inspiration from how pathologists perform diagnoses, we sought to model the whole-slide image (WSI) diagnostic problem as a decision-making one using reinforcement learning (RL). In a typical RL problem, two key components are involved: an environment and an agent, as illustrated in Figure 13. The agent starts in a given state performs an action causing the environment to transition dynamically to a new state, while receiving a reward. Our first step was to bring this interaction-driven dynamic to histopathology by developing a modular and general RL environment tailored to managing WSIs, termed HistoRL. This environment is designed to support a wide range of WSI diagnostic applications, making it a versatile framework for implementing and testing various applications. We created a toy environment based on our earlier self-supervised learning (SSL) task. This allowed us to begin the algorithm and pipeline development cycle. We trained an RL agent on several instances of this toy problem and are currently in the development phase, focusing on generalizing the agent to perform well across unseen environments during evaluation. Figure 13 shows some preliminary results. A paper has been accepted and will be presented at BioImagining 2025 conference (Porto Portugal, 2025 February 20-22).

7.15 Linear structure unfolding: application to spatial transcriptomic analysis

Participants: Morgane Fierville, Xavier Descombes, Pascal Barbry [IPMC, Sophia Antipolis], Kevin Lebrigand [IPMC, Sophia Antipolis].

Spatial transcriptomics allow serial analyses of the expression of hundreds of genes in tissue sections by a combination of molecular biology and imaging approaches. This technique allows the exploration of large tissue sections (up to 3 cm${}^2$) at the single-cell resolution, allowing cell-cell interactions and subcellular RNA expression analyses. We are focusing on developing new numerical methods to analyze spatial transcriptomics data, in particular on the Merscope and Xenium machines hosted at the IPMC. The study model in Pascal Barbry's study is the human lung and, in this project specifically, we have spatial data from biopsy of COPD (Chronic Obstrusive Pulmonary Disease) patients. In these samples, we are interested to compare smoker versus non-smoker patient and especially to follow the evolution of some cell types in specific regions such as the epithelial region (which is the first contact with cigarette smoke). But, regions of interest having particular shapes raise specific computational problems that need to be addressed (see Figure 14, top b-d). This year we have developed a method for the analysis of curved shape or point cloud (cells or RNA molecules) spread within a linear shape (see Figure 14, bottom). This method aims to unfold the corresponding shape and analyze its composition along its principal axes, in order to study cell type proportions or gene expressions. The method allows to automatically compute a centerline of the shape by combining k-means clustering and salesman problem optimization. Moreover, our method can be used iteratively to examine a shape along an orthogonal axis of the central axis (see Figure 14, top d). Therefore we are able to project transcripts into a common reference shape and study their repartition at a population level.

7.16 Detection and Segmentation of Fibronectin Structures in the Tumor Extracellular Matrix

Participants: Faisal Jayousi, Xavier Descombes, Laure Blanc-Féraud, Ellen Van Obberghen-Schilling, Emmanuel Bouilhol, Anne Sudaka [CAL, Nice].

This work is about the development of a computational pipeline for the automated detection and characterization of Fibronectin (FN) in tumor tissue on fluorescence images. FN plays a significant, yet under-explored role in tumor biology. Understanding its geometric and topological properties could provide valuable insights into cancer patients’ level of resistance to immunotherapy. The objective of this work is to design a computational pipeline for the automated detection of fibrillar FN and to propose discriminant statistical measures capable of characterizing its geometry and topology. We worked on FN from normal mouse embryo cultured fibroblasts (see Figure 15(a)) and cells treated with TGF-β1 to mimic the tumor-like phenotype of carcinoma-associated fibroblasts. From the fluorescent images of FN, a pipeline is developed to construct a graph representing the FN on which quantitative measures repesenting the geometry of the FN are computed. Statisitcs computed on 360 images show that the most discriminant features are alignment and branch length.

Future efforts will focus on adapting the pipeline for clinical immunofluorescence images, addressing challenges such as increased heterogeneity and artifacts. The open-source implementation of the pipeline (FibreSight) is available on GitHub.

7.17 Detecting symmetry in ascidian embryo

Participants: Grégoire Malandain, Patrick Lemaire [CRBM, Montpellier].

Ascidians are an appealing model to study embryogenesis since the late 19th century: they are close relative to vertebrates, some species have transparent embryos which makes them suitable for observation, and, more remarkably, embryos have a stereotyped development during the first stages, even across species. This stereotypy allowed to aggregate information from a population and compare development between individuals, thanks to a cell nomenclature 23.

For each newly acquired embryo, this requires to name its cells. This can be done by name transfer, assuming cells can be paired between two different segmented embryos. Embryos co-registration can achieve efficiently this pairing. However, the tesselated structure of embryos impairs the ability of iterative registration methods to find the sought pairing if the initial pose is out of the right convergence basin.

Exploring a (discretized) space of 3D rigid transformations is surely a way not to miss the sought transformation, but at a high computational cost. Alternatively we can take advantage of the symmetrical nature of the embryos, by aligning the symmetry planes and then only discretize the set of 2D rotations: this dramatically reduces the number of initial transformations to be tested.

In 14, we demonstrate that the cell surface contacts can define a quasi-unambiguous cell signature, and thus allows to recognize pairs of symmetrical cells, which, in turn, allow to characterize the embryo symmetry plane.

7.18 Population study of ascidian embryos

Participants: Haydar Jammoul, Grégoire Malandain, Kilian Biasuz [CRBM, Montpellier], Patrick Lemaire [CRBM, Montpellier].

During the embryogenesis, ascidian exhibit a highly stereotyped development 22. In addition, their developmental speed (rate of cell count increase) is highly consistent across embryos, up to a linear normalization (Figure 16), even if the cell division order is different between any two embryos. It suggests that the ascidian development is highly canalized, and may be simulated by a few simple rules. Eg, drawing cell lifetimes from empirical distributions allows to mimic qualitatively the number of cells of developing embryos. Our first results demonstrate that the development is governed by the cell lifetime. Additionaly, we demonstrated that, after cell division, the lifetimes of the sister cells are correlated, which was not described in the literature.

8.1 International initiatives

8.2 International research visitors

8.3 National initiatives

8.4 Regional initiatives

9.1 Promoting scientific activities

9.2 Teaching - Supervision - Juries

10.1 Major publications

• 1 articleA.Arne Bechensteen, L.Laure Blanc-Féraud and G.Gilles Aubert. A continuous relaxation of the constrained $L_2-L_0$ problem.Journal of Mathematical Imaging and VisionJanuary 2021HAL
• 2 articleM.Mayeul Cachia, V.Vasiliki Stergiopoulou, L.Luca Calatroni, S.Sébastien Schaub and L.Laure Blanc-Féraud. Fluorescence image deconvolution microscopy via generative adversarial learning (FluoGAN).Inverse Problems395April 2023, 054006HALDOIback to text
• 3 articleF.Fabienne De Graeve, E.Eric Debreuve, S.Somia Rahmoun, S.Szilvia Ecsedi, A.Alia Bahri, A.Arnaud Hubstenberger, X.Xavier Descombes and F.Florence Besse. Detecting and quantifying stress granules in tissues of multicellular organisms with the Obj.MPP analysis tool.TrafficJuly 2019HALDOI
• 4 articleX.Xavier Descombes. Multiple objects detection in biological images using a marked point process framework.Methods2016HALDOI
• 5 articleG.Georgios Efthymiou, A.Agata Radwanska, A.-I.Anca-Ioana Grapa, S.Stéphanie Beghelli-de la Forest Divonne, D.Dominique Grall, S.Sébastien Schaub, M.Maurice Hattab, S.Sabrina Pisano, M.M. Poët, D. F.Didier F Pisani, L.Laurent COUNILLON, X.Xavier Descombes, L.Laure Blanc-Féraud and E.Ellen Van Obberghen-Schilling. Fibronectin Extra Domains tune cellular responses and confer topographically distinct features to fibril networks.Journal of Cell ScienceFebruary 2021HAL
• 6 articleL.Léo Guignard, U.-M.Ulla-Maj Fiuza, B.Bruno Leggio, J.Julien Laussu, E.Emmanuel Faure, G.Gaël Michelin, K.Kilian Biasuz, L.Lars Hufnagel, G.Grégoire Malandain, C.Christophe Godin and P.Patrick Lemaire. Contact area-dependent cell communication and the morphological invariance of ascidian embryogenesis.ScienceJuly 2020HALDOI

10.2 Publications of the year

10.3 Cited publications

• 23 bookE. G.Edwin Grant Conklin. The organization and cell-lineage of the ascidian egg / by Edwin G. Conklin..Philadelphia :[Academy of Natural Sciences],1905, 182URL: http://www.biodiversitylibrary.org/item/24005back to text