Synthetic Data and its Utility in Pathology and Laboratory Medicine

Laboratory Investigation, Год журнала: 2024, Номер 104(8), С. 102095 - 102095

Опубликована: Июнь 24, 2024

Язык: Английский

---

AnyStar: Domain randomized universal star-convex 3D instance segmentation

2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Год журнала: 2024, Номер unknown, С. 7578 - 7588

Опубликована: Янв. 3, 2024

Star-convex shapes arise across bio-microscopy and radiology in the form of nuclei, nodules, metastases, other units. Existing instance segmentation networks for such structures train on densely labeled instances each dataset, which requires substantial often impractical manual annotation effort. Further, significant reengineering or finetuning is needed when presented with new datasets imaging modalities due to changes contrast, shape, orientation, resolution, density. We present AnyStar, a domain-randomized generative model that simulates synthetic training data blob-like objects randomized appearance, environments, physics general-purpose star-convex networks. As result, trained using our do not require annotated images from un-seen datasets. A single network synthesized accurately 3D segments C. elegans P. dumerilii nuclei fluorescence microscopy, mouse cortical μCT, zebrafish brain EM, placental cotyledons human fetal MRI, all without any retraining, finetuning, transfer learning, domain adaptation. Code available at https://github.com/neel-dey/AnyStar.

Язык: Английский

---

Improving 3D deep learning segmentation with biophysically motivated cell synthesis

Communications Biology, Год журнала: 2025, Номер 8(1)

Опубликована: Янв. 11, 2025

Abstract Biomedical research increasingly relies on three-dimensional (3D) cell culture models and artificial-intelligence-based analysis can potentially facilitate a detailed accurate feature extraction single-cell level. However, this requires for precise segmentation of 3D datasets, which in turn demands high-quality ground truth training. Manual annotation, the gold standard data, is too time-consuming thus not feasible generation large training datasets. To address this, we present framework generating integrates biophysical modeling realistic shape alignment. Our approach allows silico coherent membrane nuclei signals, that enable utilizing both channels improved performance. Furthermore, generative adversarial network (GAN) scheme generates only image data but also matching labels. Quantitative evaluation shows superior performance motivated synthetic even outperforming manual annotation pretrained models. This underscores potential incorporating enhancing quality.

Язык: Английский

---

A general algorithm for consensus 3D cell segmentation from 2D segmented stacks

bioRxiv (Cold Spring Harbor Laboratory), Год журнала: 2024, Номер unknown

Опубликована: Май 6, 2024

Cell segmentation is the fundamental task. Only by segmenting, can we define quantitative spatial unit for collecting measurements to draw biological conclusions. Deep learning has revolutionized 2D cell segmentation, enabling generalized solutions across types and imaging modalities. This been driven ease of scaling up image acquisition, annotation computation. However 3D which requires dense slices still poses significant challenges. Labelling every in slice prohibitive. Moreover it ambiguous, necessitating cross-referencing with other orthoviews. Lastly, there limited ability unambiguously record visualize 1000's annotated cells. Here develop a theory toolbox, u-Segment3D 2D-to-3D compatible any method. Given optimal segmentations, generates without data training, as demonstrated on 11 real life datasets, >70,000 cells, spanning single aggregates tissue.

Язык: Английский

---

Distributed and networked analysis of volumetric image data for remote collaboration of microscopy image analysis

Journal of Medical Imaging, Год журнала: 2025, Номер 12(02)

Опубликована: Март 11, 2025

The advancement of high-content optical microscopy has enabled the acquisition very large three-dimensional (3D) image datasets. analysis these volumes requires more computational resources than a biologist may have access to in typical desktop or laptop computers. This is especially true if machine learning tools are being used for analysis. With increased amount data and complexity, there need accessible, easy-to-use, efficient network-based 3D processing system. distributed networked volumetric (DINAVID) system was developed enable remote images biologists. We present an overview DINAVID compare it other currently available designed using open-source two main sub-systems, visualization with simple web interface that allows biologists upload visualization. enables model center hosting users analyzing those volumes, without manage any resources. system, tools, analyze visualize remotely also provides several including pre-processing segmentation models.

Язык: Английский

---

Improved senescent cell segmentation on bright‐field microscopy images exploiting representation level contrastive learning

International Journal of Imaging Systems and Technology, Год журнала: 2024, Номер 34(2)

Опубликована: Март 1, 2024

Abstract Mesenchymal stem cells (MSCs) are stromal which have multi‐lineage differentiation and self‐renewal potentials. Accurate estimation of total number senescent in MSCs is crucial for clinical applications. Traditional manual cell counting using an optical bright‐field microscope time‐consuming needs expert operator. In this study, the senescence were segmented counted automatically by deep learning algorithms. However, well‐performing algorithms require large numbers labeled datasets. The labeling time consuming expert. This makes learning‐based automated process impractically expensive. To address challenge, self‐supervised based approach was implemented. incorporates representation level contrastive component into instance segmentation algorithm efficient with limited data. Test results showed that proposed model improves mean average precision recall downstream task 8.3% 3.4% compared to original model.

Язык: Английский

---

Proximity adjusted centroid mapping for accurate detection of nuclei in dense 3D cell systems

Computers in Biology and Medicine, Год журнала: 2024, Номер 185, С. 109561 - 109561

Опубликована: Дек. 17, 2024

In the past decade, deep learning algorithms have surpassed performance of many conventional image segmentation pipelines. Powerful models are now available for segmenting cells and nuclei in diverse 2D types, but 3D cell systems remains challenging due to high density, heterogenous resolution contrast across volume, difficulty generating reliable sufficient ground truth data model training. Reasoning that most processing applications rely on nuclear do not necessarily require an accurate delineation their shapes, we implemented Proximity Adjusted Centroid MAPping (PAC-MAP), a U-net based method predicts position centroids proximity other nuclei. We show our outperforms existing methods, predominantly by boosting recall, especially conditions density. When trained from scratch with limited expert annotations (30 images), PAC-MAP attained average F1 score 0.793 centroid prediction dense spheroids. pretraining using weakly supervised bulk (>2300 images) followed finetuning annotations, could be significantly improved 0.816. demonstrate utility quantifying absolute content spheroids comprehensively mapping infiltration pattern patient-derived glioblastoma cerebral organoids.

Язык: Английский

---

3D Nuclei Segmentation by Combining GAN Based Image Synthesis and Existing 3D Manual Annotations

Proceedings of the 15th International Joint Conference on Biomedical Engineering Systems and Technologies, Год журнала: 2024, Номер unknown, С. 265 - 272

Опубликована: Янв. 1, 2024

Nuclei segmentation is an important task in cell analysis that requires accurate and reliable methods.In this context, deep learning based methods such as Stardist have emerged the best performing solutions for segmenting nucleus.Unfortunately, using them 3D life scientists to create new hand annotated data, a tedious especially presence of crowded population with overlapping nuclei.In work, we present workflow segment nuclei when no specific ground truth exists.Our composed three steps: first, use pre-trained 2D model every frame microscopy volume.We then train conditional GAN these paired mask frames transfer style masks.This used generate fluorescence volumes from existing data.Finally, synthetic masks.We show strategy allows data available truth, improving results obtained by training original data.

Язык: Английский

---

Ensemble processing and convexity measure for abnormally shaped nuclei segmentation

Medical Imaging 2022: Image Processing, Год журнала: 2024, Номер unknown, С. 66 - 66

Опубликована: Апрель 2, 2024

Morphological abnormalities in biological cell nuclei are used as essential features for diagnosing diseases, determining cycle stages, and conducting other fundamental research. While many deep learning approaches have been proposed segmenting normal elliptical nuclei, less work has done on abnormally shaped nuclei. One issue is that acquiring a significant number of annotated data poses challenge segmentation methods, particularly due to the generally high cost associated with obtaining data. The lack use shape analysis another problem causes not perform well To address these problems, we propose system segment limited training We generate synthetic ground truth images supplement amount available. Six Mask R-CNNs trained then introduce an ensemble strategy, known Weighted Fusion, combine results from six R-CNNs. describe step, based convexity measure, segmented result further improve performance. Our compared methods evaluation demonstrates effectiveness processing, fusion, measures

Язык: Английский

---

UNETRIS: transformer-based nuclear instance segmentation for three-dimensional fluorescence microscopy images

Medical Imaging 2022: Image Processing, Год журнала: 2024, Номер unknown, С. 62 - 62

Опубликована: Апрель 2, 2024

Automated cellular nuclei segmentation is often an important step for digital pathology and other analyses such as computer aided diagnosis. Most existing machine learning methods microscopy image analysis require postprocessing watershed transform or connected component to obtain instance from semantic results. This becomes prohibitively expensive computationally especially when used with 3D volumes. UNet Transformers Instance Segmentation (UNETRIS) proposed eliminate the steps necessary in images. UNETRIS, extension of UNETR which utilizes a transformer encoder successful "U-shaped" network design encoder-decoder structure U-Net, uses additional transformers separate individual instances cell directly during inference without need steps. UNETRIS does not but can use manual ground truth annotations training. was tested on variety volumes collected multiple regions organ tissues.

Язык: Английский

---