Clément Cazorla

Dr. Clément Cazorla

orcid

Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
1 August 2025 in Machine Learning for Biomedical Imaging
Verified
Journal Article
Open access
PDF file
CCClément Cazorla
Clément Cazorla
NMNathanaël Munier
Nathanaël Munier
RMRenaud Morin
Renaud Morin
PWPierre Weiss
Pierre Weiss
The most popular networks used for cell segmentation (e.g. Cellpose, Stardist, HoverNet,...) rely on a prediction of a distance map. It yields unprecedented accuracy but hinges on fully annotated datasets. This is a serious limitation to generate training sets and perform transfer learning. In this paper, we propose a method that still relies on the distance map and handles partially annotated objects. We evaluate the performance of the proposed approach in the contexts of frugal learning, transfer learning and regular learning on regular databases. Our experiments show that it can lead to substantial savings in time and resources without sacrificing segmentation quality. The proposed algorithm is embedded in a user-friendly Napari plugin.

21 May 2024 in Scientific Reports
Verified
Journal Article
Open access
PDF file
CCClément Cazorla
Clément Cazorla
RMRenaud Morin
Renaud Morin
PWPierre Weiss
Pierre Weiss
We present Svetlana (SuperVised sEgmenTation cLAssifier for NapAri), an open-source Napari plugin dedicated to the manual or automatic classification of segmentation results. A few recent software tools have made it possible to automatically segment complex 2D and 3D objects such as cells in biology with unrivaled performance. However, the subsequent analysis of the results is oftentimes inaccessible to non-specialists. The Svetlana plugin aims at going one step further, by allowing end-users to label the segmented objects and to pick, train and run arbitrary neural network classifiers. The resulting network can then be used for the quantitative analysis of biophysical phenoma. We showcase its performance through challenging problems in 2D and 3D and provide a comprehensive discussion on its strengths and limits.

8 April 2026
Verified
Thesis
Closed access
CCClément Cazorla
Clément Cazorla
Dans ce manuscrit de thèse, nous explorons l'usage de réseaux de neurones dans un contexte frugal, c'est-à-dire avec peu de données d'entraînement, appliqué à des tâches d'analyse d'images de microscopie. Au cours de ces travaux, j'ai fourni un travail substantiel d'intégration dans des interfaces utilisateur ergonomiques. Celles-ci prennent la forme de plugins Napari, un logiciel open-source de visualisation d'images multi-dimensionnelles. Plus largement, ce travail permet de rapprocher les utilisateurs de microscopes et les technologies d'apprentissage profond, actuellement complexes à appréhender, et à implémenter, à travers des logiciels d'apprentissage interactifs. Les biologistes recourent de plus en plus à diverses modalités d'imagerie, dans le but d'accéder à de nouvelles sources d'information dans des contextes de recherche fondamentale comme applicative. Cette intensification de l'usage des microscopes a conduit à des volumes de données importants, dont l'analyse pose aujourd'hui de réelles questions techniques et scientifiques. Le traitement de ces données est bien souvent effectué de manière manuelle par des biologistes (comptage de cellules par exemple), induisant des problèmes de reproductibilité des résultats inter et intra opérateurs, mais aussi des temps de traitement parfois très importants. L'essor de l'apprentissage profond au cours des dernières années donne aujourd'hui accès à des possibilités d'analyse encore inatteignables il y a quelques années de cela. Néanmoins, à l'heure où les questions de consommation d'énergie et de stockage deviennent cruciales, ces approches posent certains problèmes. Le recueil de grandes quantités d'images et leur annotation sont coûteux et complexes. Leur fiabilité est parfois remise en cause, car ces modèles neuronaux sont fréquemment considérés comme difficiles, voire impossibles à interpréter. À l'heure actuelle, de nombreuses initiatives de mise en commun de modèles et de ressources de calculs, ainsi que de recherche sur l'interprétabilité de ces réseaux sont déployées, afin de les rendre accessibles à un plus large public. Cette thèse s'articule autour de deux grands axes de l'analyse d'images que sont la segmentation et la classification. Dans un premier temps, nous rappellerons l'état de l'art des méthodes d'analyse par apprentissage en microscopie, puis nous présenterons les méthodes frugales que nous avons développées pour ces deux tâches. Par la suite, nous présenterons leur intégration sous forme de plugins Napari open-source. Enfin, nous montrerons, à travers des exemples concrets de collaborations à la fois académiques et industrielles, l'intérêt de ces méthodes et les résultats qu'elles fournissent.

17 February 2025 in bioRxiv
Verified
Preprint
Open access
PDF file
LBLair Benjamin
Lair Benjamin
CCClément Cazorla
Clément Cazorla
ALAlicia Lobeto
Alicia Lobeto
ALAxel Labour
Axel Labour
CLClaire Laurens
Claire Laurens
CMCedric Moro
Cedric Moro
Background The myogenic fusion index (FI) is commonly used in skeletal muscle cell culture to assess the ability of myoblasts to form myotubes, as the ratio of myoblast nuclei fused with myotubes over the total number of myoblasts. The manual quantification of the FI from 2D microscopy images is tedious and biased, thus several automated methods have been developed. However, they still face challenges such as efficient nucleus segmentation and classification of fused and isolated myoblast nuclei. Here, we developed a novel workflow entirely based on AI for fully automated and unbiased quantification of the FI. Results Using current methods, we show that myoblast nuclei located above or below myotubes can significantly corrupt accurate FI computation. To circumvent this issue, we developed MyoFuse which enables an accurate and high-throughput segmentation and classification of myonuclei. It comprises a nuclei segmentation step using Cellpose, followed by a classification network trained with Svetlana. MyoFuse demonstrated strong accuracy when tested against manual annotation in mouse C2C12 and human primary myotubes. The trained classifier is able to differentiate myotube nuclei from myoblast nuclei based on myotube cytoplasm staining only. Experimental comparisons also highlighted that the previously developed methods lead to a significant overestimation of the FI. Conclusion In summary, we underscore the lack of accuracy of traditional methods for automated FI quantification. MyoFuse enables a direct and accurate segmentation of nuclei even in nuclei clusters frequently observed in myotubes. This workflow thus offers a new and more reliable method to evaluate the FI. It also limits the selection bias by processing large images.

8 December 2023 in HAL
Verified
Preprint
Open access
PDF file
CCClément Cazorla
Clément Cazorla
NMNathanaël Munier
Nathanaël Munier
Renaud MorinRenaud Morin
Renaud Morin
PWPierre Weiss
Pierre Weiss
A few neural networks in biological image segmentation rely on a prediction of a distance map. This principle is at the basis of popular software such as Stardist, Cellpose or Omnipose. It yields unprecedented accuracy but hinges on fully annotated datasets. This can be a serious limitation for generating training sets and performing transfer learning. In this paper, we show how to handle partial annotation, while still relying on the distance map. We design a variant of the Omnipose architecture embedded in a user-friendly Napari plugin. We evaluate the performance of the proposed approach in the contexts of frugal learning, transfer learning and regular learning on a large database. Our experiments show that the proposed approach can lead to substantial savings in time and resources without sacrificing segmentation quality.

25 August 2025 in arXiv
Verified
Preprint
Open access
PDF file
CCClément Cazorla
Clément Cazorla
NMNathanaël Munier
Nathanaël Munier
RMRenaud Morin
Renaud Morin
PWPierre Weiss
Pierre Weiss
The most popular networks used for cell segmentation (e.g. Cellpose, Stardist, HoverNet,...) rely on a prediction of a distance map. It yields unprecedented accuracy but hinges on fully annotated datasets. This is a serious limitation to generate training sets and perform transfer learning. In this paper, we propose a method that still relies on the distance map and handles partially annotated objects. We evaluate the performance of the proposed approach in the contexts of frugal learning, transfer learning and regular learning on regular databases. Our experiments show that it can lead to substantial savings in time and resources without sacrificing segmentation quality. The proposed algorithm is embedded in a user-friendly Napari plugin.