Fig. 2

Training and Benchmarking CellSeg performance on the 2018 Kaggle data challenge. A Information on Kaggle dataset used to develop, train, and test CellSeg. CellSeg final performance was assessed on a test set provided by the Kaggle data challenge using mean average precision (mAP) score. B CellSeg segmentation of representative fluorescence image from the Kaggle test set. White arrowheads: cells with blurred nuclear boundaries C CellSeg segmentation of representative H&E-stained brightfield image. Red arrows: nuclear debris. D CellSeg performance compared to other top performing segmentation algorithms in data science bowl. Columns show mean average precision (mean AP) scores reported on Kaggle DSB2018 stage 2 test set and average F1 scores. For nucleAIzer, reported scores from the original publication [29] are displayed. For StarDist, brightfield and fluorescence images were segmented using 2D_versatile_he pre-trained model and 2D_versatile_fluo pre-trained model, respectively. For Cellpose, the pre-trained nuclei segmentation model was used (see “Methods” section for testing details)