Cerebral organoids have become a fundamental biological model system across developmental neuroscience, offering insights into early human brain development, neurological diseases and drug response in vitro. Despite its promise, cerebral organoid research remains constrained by the limitations of current analytical methods for organoid imaging and characterisation, which are destructive, require exogenous labelling and have limited multiplexing capabilities. Here, we present a non-invasive, label-free organoid imaging approach based on Raman microspectroscopy combined with physics-constrained autoencoder representation learning. Our approach enables in situ biochemical imaging of intact cerebral organoids, allowing the interrogation of structural, biochemical and morphological composition at cellular and near-subcellular resolution without compromising organoid integrity.