Overcoming segmentation bottlenecks in brain tissue for spatial proteomic inference at population scale
Abstract
Tissue imaging depends on cell segmentation, whose errors carry into downstream analyses. In human brain, pretrained models struggle with complex cell bodies, and the segmentation accuracy has rarely been evaluated. Here, we introduce SPInfer, a workflow developed using thousands of expert-annotated cell bodies from 120 immunofluorescence images of human prefrontal cortex. After benchmarking eight methods, we fine-tuned the best-performing foundation model, thereby more than doubling cell segmentation accuracy. The resulting cell compositions agreed with single-nucleus RNA-seq and spatial transcriptomics. Across 14,765 images from 69 donors, zero-shot masks showed no significant association with Alzheimer’s disease pathology, whereas fine-tuned masks identified four; three of these linked OLIG2-positive cell fractions to tangle and plaque burden. Fine-tuned masks also enabled cell typing from the nuclear stain alone and detected genes associated with cell morphology. This work provides a general pipeline to segment complex tissues, enabling biological discovery in spatial proteomics at the population scale.