The study team developed SPInfer, a segmentation-to-inference pipeline designed for large-scale spatial proteomics in human brain tissue. They assembled a manually labeled training corpus comprised of 5,832 cell-body annotations drawn from 120 immunofluorescence images of human prefrontal cortex. These human annotations served as the reference standard for benchmarking and for adapting pretrained models to the morphological complexity of cortical cells. The authors reported that pretrained models struggled on brain tissue; the manual labels were therefore central to adapting a foundation model for reliable downstream quantification.
Investigators compared eight segmentation approaches spanning classical algorithms, convolutional neural networks, and foundation-model-based methods. They selected Cellpose-SAM for adaptation based on benchmarking results. After fine-tuning on the annotated corpus, reported cell segmentation F1 rose from 0.37 using zero-shot masks to 0.84 with the fine-tuned model. The study team observed that this performance lift produced cell compositions that aligned with matched single-nucleus RNA-seq and spatial transcriptomics measurements in their dataset. The source did not enumerate all eight tools by name in the abstract, nor did it provide detailed per-class performance metrics within the abstract text; those specifics were not reported there.
The authors applied the fine-tuned pipeline across a population cohort: 14,765 immunofluorescence images from 69 postmortem donors. This scale tested whether improved segmentation would translate into different downstream biological inferences when compared with zero-shot masks from a pretrained foundation model. The workflow processed tens of thousands of images and generated per-image and per-nucleus quantifications intended for association testing with pathology measures.
The study team directly compared association results produced using zero-shot masks versus fine-tuned masks. When investigators used zero-shot masks, identical models of Alzheimer's disease pathology returned no significant associations after false discovery rate correction. By contrast, using the fine-tuned Cellpose-SAM masks yielded four significant associations after FDR correction. Three of those linked higher fractions of OLIG2-positive cells to increased neurofibrillary tangle and plaque burden. The authors therefore reported that segmentation quality materially affected whether pathology associations were discovered in these spatial proteomic data.
With fine-tuned masks in hand, investigators trained a graph attention network over nuclear neighborhoods. The network classified four cell types without using canonical marker information, according to the abstract. In addition, differential expression analyses tied morphological variation to gene expression patterns in the dataset. The abstract did not provide class labels for the four cell types, nor did it report specific genes implicated in the morphology–expression links; those details were not included in the abstract text and therefore are not stated here.
The authors have shared resources: a Hugging Face dataset and a GitHub repository were listed in the article footnotes as publicly available resources. The preprint has not been certified by peer review. The abstract explicitly states that prior pretrained segmentation models struggled with morphologically complex brain cell bodies and that "how domain adaptation affects downstream discovery remains unclear." That question — whether and how segmentation domain adaptation systematically alters biological discovery in spatial proteomics across cohorts — remains an open question documented by the authors. The abstract also does not report certain experimental specifics in its text (for example, the full list of benchmarked segmentation tools, per-class confusion matrices, or exact network hyperparameters), and those methodological details were not reported there.
Limitations and next clarifications therefore center on replication and transparency: how generalizable the fine-tuned model is across other brain regions, staining panels, and imaging platforms; the sensitivity of downstream associations to alternative segmentation strategies; and independent validation through peer review. These items reflect outstanding questions that the authors themselves framed in the preprint.