Light Sheet Microscopy Foundation Model Pretraining Dataset

This dataset consists of 3D image patches extracted from light sheet microscopy (LSM) volumes with paired text descriptions. Images are drawn from five sources: an internal dataset from the Wu Lab at Weill Cornell Medicine (New York, NY, USA), and four public datasets: SELMA3D MICCAI challenge, Allen human brain, Allen developing mouse brain, and Allen viral labeling of projection neurons. From each dataset, up to 10 non-overlapping 96×96×96 voxel patches were randomly selected with a minimum foreground threshold of 5%, yielding 1,023 patches in total. The patches capture diverse neuroanatomical structures and imaging conditions across species and brain regions. Text descriptions were authored by a domain expert and paraphrased using an LLM to increase linguistic diversity while preserving semantic fidelity. This dataset was used for pretraining a multimodal LSM foundation model.