Spheroid LSFM dataset with nuclei centroid annotations

Description This repository contains 1) A dataset of lightsheet images of spheroids of two different cell lines (LN18-RED and SH-SY5Y) and centroid annotations for model training and evaluation; and 2) Weights of nuclei centroid prediction models trained on this data. See the pre-printreferenced below for more information.If you use this dataset or the models in your research, please cite our paper: Van De Looverbosch, Tim, Sarah De Beuckeleer, Frederik De Smet, Jan Sijbers, and Winnok H. De Vos. “Proximity Adjusted Centroid Mapping for Accurate Detection of Nuclei in Dense 3D Cell Systems.” Computers in Biology and Medicine 185 (February 1, 2025): 109561. https://doi.org/10.1016/j.compbiomed.2024.109561. The code accompanying our paper is available at https://github.com/DeVosLab/PAC-MAP. Dataset Organization The dataset is organized in the following way: LN18-RED pretrain (pretraining data with weak annotations) patches sample_01 patch_0000.tif ... ... patches_binary (used to filter out patches with a lot of background) sample_01 patch_0000.tif ... ... points_csv (weak nuclei centroid annotations from seeded watershed) sample_01 patch_0000.csv ... ... SH-SY5Y train (training/validation data with manual annotations) patches sample_XX patch_xxxx.tif ... ... points_csv sample_XX patch_xxxx.csv ... ... test (test data with manual anntotations) patches sample_XX patch_yyyy.tif ... ... points_csv sample_XX patch_yyyy.csv ... ... preprocessed_images (preprocessed in-toto images added for reference) LN-18-RED sample_01.tif ... SH-SY5Y sample_01.tif ... Dataset generation See the materials and methods in our pre-print. Models Pretrained, trained from scratch and finetuned models (C-MAP and PAC-MAP versions) for nuclei centroid prediction are provided for this LSFM spheroid dataset and for the LSFM mouse brain dataset provided by Krupa et al. (2021) (https://doi.org/10.1016/j.celrep.2021.109802). See our pre-print for usage, model training procedures and model performance. For each model type (pretrained, scratch, finetuned), model training was repeated three times with a different random seed (indicated in the model name).