Global morphological heterogeneity of retinal organoids across development. Top row: Image segmentation and analysis pipeline. A convolutional neural network (CNN) with the DeepLabV3 architecture was trained on 841 images and their manually annotated masks (left). The CNN was subsequently used to segment all images of the dataset that were finally subjected to a pipeline extracting a total of 165 morphological parameters (morphometrics, right), including, among others, shape descriptors and image moments. Intra-experimental global morphological organoid heterogeneity. Time-series images from one representative experiment were analyzed using the image analysis pipeline and subjected to t-SNE dimensionality reduction on the first 20 principal components. Data points were colored by individual organoids (left graph) and time frames of organoid development within the imaging window (right graph). While organoids clustered closely at earlier time points (up to 24 h), they strongly diverged at later time points, suggesting increasing inter-individual changes of their morphological characteristics over time.
Organoids are key models for studying development & disease, but heterogeneity is a problem. @wittbrodtlab.bsky.social use #DeepLearning to predict differentiation paths & resulting tissues in #retinal organoids, with implications for other #organoid systems @plosborn.bsky.social 🧪 plos.io/45VTMt1