Tellu – an object-detector algorithm for automatic classification of intestinal organoids

Abstract

Intestinal epithelial organoids recapitulate many of the in vivo features of the intestinal epithelium, thus representing excellent research models. Morphology of the organoids based on light microscope images is used as a proxy to assess the biological state of the intestinal epithelium. Currently, organoid classification is manual and therefore subjective and time-consuming, hampering large-scale quantitative analyses. Here we describe Tellu, an object detector algorithm trained to classify cultured intestinal organoids. Tellu was trained by manual annotation of >20000 intestinal organoids to identify cystic non-budding organoids, early organoids, late organoids, and spheroids. Tellu can also be used to quantify relative organoid size. Tellu classifies intestinal organoids to these four subclasses with accuracy comparable to trained scientists but significantly faster and without bias. Tellu is provided as an open, user-friendly online tool to benefit the increasing number of investigations using organoids through fast and unbiased organoid morphology and size analysis.

Identification of a targetable ST2-expressing fibroblast subset driving Peutz-Jeghers syndrome polyposis

“…Intestinal epithelial organoids grow in Matrigel as three-dimensional structures recapitulating the homeostatic crypt-villus architecture of intestinal epithelium (Sato et al, 2009), but tumorigenic and regenerative organoids fail to form the correct 3D-structure and instead grow as hyperproliferative spheroids (Langlands et al, 2018; Nusse et al, 2018). As determined by eye and quantified using our automated organoid classification tool “Tellu” (Domènech-Moreno et al, 2023), Lkb1 -deficient fibroblasts induced an increased frequency of spheroids as compared to organoids in the cocultures, indicating that a regenerative/tumorigenic phenotype of the WT intestinal epithelium is promoted by Lkb1 -deficient fibroblasts (Figure 4C). This was consistent with Lkb1 -deficient fibroblasts promoting polyp formation in vivo .…”

Suppression of epithelial proliferation and tumorigenesis by immunoglobulin A

“…Imaging was performed for 4-5 days after adding treatments using Z-stack on a Keyence BZ-X800 epifluorescence microscope. Image quantifications were made in ImageJ 2.14.0 and Tellu 76 , and all image calculations were made in R studio 4.2.2.…”

A Curated Cell Life Imaging Dataset of Immune-enriched Pancreatic Cancer Organoids with Pre-trained AI Models

“…Up to now, several algorithms have emerged to analyze the response of organoids during live cell imaging to specific treatments based on physical measurements of organoid properties, such as mean area, organoid eccentricity and total organoid count. These algorithms enable the continuous assessment of organoid reactions over time without disrupting the organoid system [14][15][16][17][18] . Nevertheless, each algorithm is evaluated on its proper dataset, in optimal conditions, not allowing a valid comparison among the different performances.…”

“…It is distinguished by meticulous manual annotations for every field of view. Its dimension, in terms of number of images, variety and number of organoids, is similar or higher than other datasets used to validate novel segmentation protocols 14,16,18 .…”

“…It is distinguished by meticulous manual annotations for every field of view. Its dimension, in terms of number of images, variety and number of organoids, is similar or higher than other datasets used to validate novel segmentation protocols 14,16,18 . The substantial variability in organoid density across the dataset images makes it a particularly challenging benchmark for the evaluation of segmentation protocols.…”