BIDCell is a self-supervised deep learning framework for segmenting subcellular spatial transcriptomics data, incorporating biologically-informed loss functions to learn relationships between spatial gene expression and cell morphology. It integrates cell-type data and morphology information to improve segmentation accuracy. BIDCell outperforms existing methods across various tissue types and platforms, demonstrating its potential to enhance single-cell spatial expression analysis. The framework uses a comprehensive evaluation system, CellSPA, to assess segmentation performance across five categories. BIDCell captures diverse cell shapes and improves expression purity, reducing contamination from neighboring cells. It is replicable and generalizable across different SST platforms. BIDCell's performance is validated on multiple datasets, including Xenium-BreastCancer, Xenium-MouseBrain, CosMx-Lung, MERSCOPE-Melanoma, and Stereo-seq-MouseEmbryo. It accurately segments cells and identifies region-specific subtypes, such as neuronal cell types. BIDCell's biologically-informed loss functions enable it to self-learn from spatial transcriptomic and DAPI images, improving segmentation accuracy without requiring manual annotations. The framework is flexible and can be applied to various technologies and gene panels. BIDCell's evaluation system, CellSPA, provides a comprehensive assessment of segmentation performance, promoting reproducible science and transparency. The method has the potential to revolutionize subcellular spatial transcriptomics by enabling more accurate cell segmentation and biological discovery.BIDCell is a self-supervised deep learning framework for segmenting subcellular spatial transcriptomics data, incorporating biologically-informed loss functions to learn relationships between spatial gene expression and cell morphology. It integrates cell-type data and morphology information to improve segmentation accuracy. BIDCell outperforms existing methods across various tissue types and platforms, demonstrating its potential to enhance single-cell spatial expression analysis. The framework uses a comprehensive evaluation system, CellSPA, to assess segmentation performance across five categories. BIDCell captures diverse cell shapes and improves expression purity, reducing contamination from neighboring cells. It is replicable and generalizable across different SST platforms. BIDCell's performance is validated on multiple datasets, including Xenium-BreastCancer, Xenium-MouseBrain, CosMx-Lung, MERSCOPE-Melanoma, and Stereo-seq-MouseEmbryo. It accurately segments cells and identifies region-specific subtypes, such as neuronal cell types. BIDCell's biologically-informed loss functions enable it to self-learn from spatial transcriptomic and DAPI images, improving segmentation accuracy without requiring manual annotations. The framework is flexible and can be applied to various technologies and gene panels. BIDCell's evaluation system, CellSPA, provides a comprehensive assessment of segmentation performance, promoting reproducible science and transparency. The method has the potential to revolutionize subcellular spatial transcriptomics by enabling more accurate cell segmentation and biological discovery.