#3 · ML / AI methods

Benchmarking segmentation and synthetic ground truth for 3D tissue and organoid imaging

The dominant thread is quantitative image analysis in tissue geometry, where segmentation accuracy and reconstruction fidelity bound what can be measured. Two items benchmark segmentation directly: CRISP compares five algorithms across ten mouse tissues on Xenium spatial transcriptomics using a ground-truth-free purity metric, and a separate evaluation tests Cellpose-SAM, microSAM, CellSAM, Mesmer and InstanSeg on live microscopy and CODEX spatial-omics data. Other items address ground truth and reconstruction: a parametric physics-based framework generates 3D organoid fluorescence images with exact cell and nucleus masks, and a volume electron microscopy pipeline segments gold nanoparticles and cells in SBF-SEM tumor spheroids. ToFiE reconstructs dense 3D fiber networks from microscopy, and LEAO improves aberration estimation more than sixfold for intravital fluorescence imaging. The new elements are the ground-truth-free metric and synthetic ground truth; the work is primarily method development and benchmarking rather than biological conclusions.