To characterize the growth of brain organoids (BOs), cultures that replicate some early physiological or pathological developments of the human brain are usually manually extracted. Due to their novelty, only small datasets of these images are available, but segmenting the organoid shape automatically with deep learning (DL) tools requires a larger number of images. Light U-Net segmentation architectures, which reduce the training time while increasing the sensitivity under small input datasets, have recently emerged. We further reduce the U-Net architecture and compare the proposed architecture (MU-Net) with U-Net and UNet-Mini on bright-field images of BOs using several data augmentation strategies. In each case, we perform leave-one-out cross-validation on 40 original and 40 synthesized images with an optimized adversarial autoencoder (AAE) or on 40 transformed images. The best results are achieved with U-Net segmentation trained on optimized augmentation. However, our novel method, MU-Net, is more robust: it achieves nearly as accurate segmentation results regardless of the dataset used for training (various AAEs or a transformation augmentation). In this study, we confirm that small datasets of BOs can be segmented with a light U-Net method almost as accurately as with the original method.

Brain organoids are three-dimensional tissues gener-ated in vitro from pluripotent stem cells and replicating the early development of Human brain. To implement, test and compare methods to follow their growth on microscopic images, a large dataset not always available is required with a trusted ground truth when developing automated Machine Learning solutions. Recently, optimized Generative Adversarial Networks prove to generate only a similar object content but not a background specific to the real acquisition modality. In this work, a small database of brain organoid bright field images, characterized by a shot noise background, is extended using the already validated AAEGAN architecture, and specific noise or a mixture noise injected in the generator. We hypothesize this noise injection could help to generate an homogeneous and similar bright-field background. To validate or invalidate our generated images we use metric calculation, and a dimensional reduction on features on original and generated images. Our result suggest that noise injection can modulate the generated image backgrounds in order to produce a more similar content as produced in the microscopic reality. A validation of these images by biological experts could augment the original dataset and allow their analysis by Deep-based solutions.

Génération, extraction et caractérisation sur des images de culture d'organoides cérébraux avec des contraintes de petits datasets: Les organoïdes cérébraux (OC) sont des modèles récents et prometteurs pour le suivi du développement cérébral. Cependant, les biologistes manquent d’outils d’analyse automatique ou semi-automatique pour suivre leur développement in vitro, du fait de la faible quantité d’images disponibles actuellement. Dans ce contexte, cette thèse vise à automatiser l’extraction et à caractériser de façon adaptée la morphologie des organoïdes cérébraux sur de petits datasets d’images acquises en microscopie à divers stades de la croissance de ces cultures.Dans cet objectif, un petit dataset a été augmenté par divers réseaux adverses génératifs (GAN) et le meilleur d’entre eux au regard de critères psychovisuels, statistiques et quantitatifs a été optimisé. Quelle que soit l’optimisation utilisée pour générer des images, elle enrichit le dataset de manière constructive, excepté pour un groupe. De manière complémentaire, l’effet d’une variation du bruit par GAN pour générer des images plus proches des exceptions a été étudié, et les images générées ont été validées également automatiquement, statistiquement et psycho-visuellement.Dans la continuité de cette première contribution, la forme des OC originels a été ex-traite par deep learning et nous montrons qu’une meilleure précision peut être obtenue sil'entraînement est réalisé par un GAN optimisé ́e. L’architecture de l’extraction a été réduite pour l’adapter aux petits datasets.Enfin, grâce aux images segmentées, la morphologie des OC a pu être caractérisée par calcul d’indices morphologiques, mais également en déterminant leur empreinte par analyse de données topologiques. Caractériser un indice de leur développement (neuroépithélium) en association avec un autre laboratoire serait une étape future comme la validation de cette approche sur d’autres cultures cérébrales et pour diverses modalités d’acquisition microscopiques.

The Digital Imaging and Communications in Medicine (DICOM) format is widely used to store medical images along with patient-related Protected Health Information (PHI) within hospital Picture Archiving and Communication System (PACS). However, the compliant anonymization of DICOM objects for PACS-to-PACS transmission remains challenging. Our objective is to identify a lightweight conditioned Generative Adversarial Network (cGAN) adapted from Scene Text Removal (STR) models, capable of effectively removing burned-in text from DICOM images while restoring the underlying tissue structures. To this aim, we generated a synthetic dataset with 620 images from The Cancer Imaging Archive (TCIA) database, containing four modalities with fake PHI. A localized generation strategy is applied to minimize unintended errors outside the burned-in text regions, as they typically occupy only a small portion. Then, five cGAN-based architectures are fine-tuned and compared to evaluate their reconstruction performance, FLOPs and inference time. After fine-tuning, all models are subsequently evaluated on CPU-only inference to assess their real-world transferability, identifying GaRNet model as the best trade-off across different evaluation aspects. Compared with the second-best cGAN, MTRNet++, GaRNet required 23.6 times fewer FLOPs and achieved 11.4 times faster inference while maintaining comparable structural similarity and PSNR score.