BARseq drops manual annotation: training on instrument-matched optics recovers about four more transcripts per cell

Transcript calling in BARseq no longer depends on manually annotated fluorescent-spot ground truth, because the detector is trained on semi-synthetic images generated from the instrument's optical model. BARseq analysis previously relied on BarDensr's single-shot density estimator, which produced comparatively conservative calls in crowded regions with overlapping signal, while supervised deep learning needs manual ground truth that is hard to obtain at scale in real tissue images. Across 10 representative fields of view, PhySeqNet and PhySeqNet with NMSB achieved the highest reconstruction fidelity by both NCC and MS-SSIM, with mean NCC of 0.862±0.024 and 0.848±0.018 and mean MS-SSIM of 0.702±0.026 and 0.706±0.025, compared by one-sided paired Wilcoxon signed-rank tests with Benjamini-Hochberg correction.

The detector treats fluorescent foci as emitters to be demixed rather than as foreground to be segmented. Classical LoG, DoG, and h-maxima detectors depend strongly on scale and intensity-threshold parameters and are not designed to infer multiple overlapping emitters from a single blended point-spread-function spot. Training images are built by convolving three-dimensional emitter volumes with a Born and Wolf confocal point-spread function and adding real sample-free experimental noise frames pixel-wise, with emitter centroids as ground truth; on held-out nonblank simulated images Model G reached precision 0.921 and recall 0.879, and Model C 0.934 and 0.882.

Barcode-aware non-maxima suppression collapses duplicate detections of the same rolony while keeping nearby detections assigned to different barcodes. Raw PhySeqNet was trained to favor dense predictions, so it can produce multiple detections corresponding to the same underlying rolony in crowded regions. Detections closer than 2 pixels (under 0.66 µm) and assigned to the same barcode were collapsed to the higher-scoring one; on the simulated test set this step removed about 40% of true positives while removing about 66% of false positives, so the post-NMSB count is an intentionally conservative estimate.

The work targets researchers and image-analysis engineers using BARseq spatial transcriptomics, and validation covers a single coronal brain section from one mouse across 60 fields of view, where the detector and the pyBARseq pipeline are demonstrated end to end. The detector is designed for BARseq-like confocal imaging, and the optical parameters, noise frames, and spot densities used for training and evaluation come from that imaging condition. The authors state that extending the physics-informed training framework to other smFISH modalities would require regenerating simulations with matched optical parameters and background and noise conditions and retraining.

Validation is performed on a single coronal brain section from one animal, and the full pipeline including cell typing and MERFISH validation is demonstrated only on that dataset. On the Piscis smFISH benchmark, images with similar optical and emitter characteristics transfer partially zero-shot, performance degrades on images with high diffuse background and very sparse spots, and it fails on near-empty fields dominated by background structure. The authors note that using sample-free experimental frames rather than a parametric noise model follows from fixed-pattern noise and instrument-specific background being difficult to capture analytically, and that a direct comparison with parametric noise models was beyond the scope of the work.