Random forest segmentation and spatial analysis of injected cardiac spheroids in optically cleared myocardium

Accurate quantification of transplanted cardiac spheroids requires three-dimensional (3D) localisation within intact myocardium, yet this remains technically challenging. Optical clearing and light-sheet microscopy enable volumetric imaging of injection sites, but automated segmentation is difficult when transplanted spheroids and host tissue are labelled with the same fluorescent markers and cannot be separated by simple intensity thresholding. We developed an application-focused workflow for 3D segmentation and spatial analysis of injected hiPSC-derived cardiomyocyte and H9c2 spheroids in optically cleared rabbit myocardium. Random forest pixel classification using intensity, edge, and texture features was used to segment spheroids, with connected-component analysis subsequently used to reconstruct individual objects. Compared with simple intensity thresholding, random forest classification provided substantially improved segmentation, although a U-Net achieved higher segmentation performance. The resulting 3D segmentations enabled automated extraction of spheroid size, nearest-neighbour distance, and epicardial depth. This workflow provides an end-to-end pipeline from volumetric light-sheet microscopy of injected cardiac spheroids to automated 3D spatial quantification, using sparse annotations to support segmentation of large cleared-tissue datasets.