Machine learning

The method proposed in this paper is a robust combination of multi-task learning and unsupervised domain adaptation for segmenting amoeboid cells in microscopy. This end-to-end framework provides a consolidated mechanism to harness the potential of multi-task learning to isolate and segment clustered cells from low contrast brightfield images, and it simultaneously leverages deep domain adaptation to segment fluorescent cells without explicit pixel-level re- annotation of the data.

The entry-point to the codebase is the main.py file. The user has the option to

• Train the network on their own dataset
• Load a pre-trained model and use that for inference on their own data
• Note: The provided pretrained model was trained on 256x256 images. Results on different resolutions could require fine-tuning This model is trained (supervised) on brightfield, and domain adapted to fluorescence data. The results are saved as 'inference.png'

This workflow describes a deep-learning based pipeline for reliable single-organoid segmentation and tracking in 2D+t high-resolution brightfield microscopy of mouse mammary epithelial organoids. The pipeline involves a four-layer U-Net to infer semantic segmentation predictions, adaptive morphological filtering to establish candidate organoid instances, and a shape-similarity-constrained, instance-segmentation-correcting tracking step to associate the corresponding organoid instances in time.

It is particularly focused on automatically detecting an organoid located approximately in the center of the first frame and track all its subsequent instances in the remaining frames, emphasizing on accurate organoid boundary delineation. Furthermore, segmentation network was trained using plausible pix2pixHD-generated bioimage data. Syntheric image simulator code and data are also available here.

OrganoID is an image analysis platform that automatically recognizes, labels, and tracks single organoids, pixel-by-pixel, in brightfield and phase-contrast microscopy experiments. The platform was trained on images of pancreatic cancer organoids and validated on separate images of pancreatic, lung, colon, and adenoid cystic carcinoma organoids.

Fractal is a framework to process high-content imaging data at scale and prepare it for interactive visualization. Fractal provides distributed workflows that convert TBs of image data into OME-Zarr files. The platform then processes the 3D image data by applying tasks like illumination correction, maximum intensity projection, 3D segmentation using cellpose and measurements using napari workflows. The pyramidal OME-Zarr files enable interactive visualization in the napari viewer.

This workflow applies a Stardist pre-trained model (versatile_fluo or versatile_HE) depending on the input images ie. uses both models for a dataset including both fluorescence (grayscale or RGB where all channels are equal) and H&E stained (RGB where channels are not equal) images.

This version uses tensorflow CPU version (See Dockerfile) to ensure compatibility with a larger number of computers. A GPU version should be possible by adapting the Dockerfile with tensorflow-gpu and/or nvidia-docker images.