Loading…
Nuclear segmentation and subtype classification in immunohistochemistry slides, steering frozen SAM and UNI encoders with frequency-domain prompts.
No providers recorded yet. Browse all providers
In a CD47-stained slide, the brown diaminobenzidine deposit is not a rendering artifact to be normalized away — it is the measurement. A nucleus counts as positive because it carries that chromogen, and what the report says depends on whether those positive nuclei are tumor or immune. Models trained on hematoxylin and eosin have no reason to read that channel, and the boundary contrast they do rely on is what a heavy DAB load washes out. Stain normalization suppresses the biomarker signal itself, and the authors report that it and cross-domain transfer have both yielded only marginal gains.
FPNuNet answers this by refusing to retrain the backbones at all. A frozen Segment Anything ViT-B supplies structural features and a frozen UNI ViT-L supplies pathology semantics; both are steered by lightweight prompt generators built on the discrete cosine transform, which is where staining intensity and nuclear boundary texture separate cleanly. Two further streams run in parallel from the raw patch — a Haar wavelet encoder and a three-scale spatial context encoder — and a DCT-enabled fusion neck combines all four before three decoder branches predict binary masks, horizontal–vertical distance maps, and nuclear type. The output structure follows HoVer-Net, the same lineage as CellViT, but the encoder half is adapted rather than fine-tuned.
The work comes from the computer science and artificial intelligence schools of Shenzhen University, with clinical collaborators at Chongqing University Jiangjin Hospital, the University of Birmingham and the University of Oxford. It was published in GigaScience in September 2026.
Patches enter at 128×128 RGB and leave as binary, boundary, HV and type logits. Training uses AdamW with a 1e-3 learning rate for the prompt generators and 5e-4 for the remaining trainable components, a 1,000-step warmup, 0.5× decays at steps 20,000 and 27,000, 30,000 total steps, and bf16 mixed precision. The loss combines BCE, Dice and focal terms on the binary branch with a 0.5× boundary term, foreground-normalized MSE plus Sobel gradient error on HV, and equally weighted cross-entropy, Dice, focal and soft-IoU on type.
CD47-IHCNuSC comprises 86 non-overlapping 1024×1024 patches cut from CD47-immunostained esophageal cancer whole-slide images, digitized at 0.25 µm/pixel on an Aperio AT2 and split 70/16 for training and test. Its 18,483 nuclei carry seven phenotype-aware subtypes — positive and negative tumor, immune, stromal and other — assigned through a human-in-the-loop pipeline of automated pre-segmentation, manual correction in QuPath, and board-certified pathologist adjudication, with no stain normalization applied. The distribution is long-tailed: positive tumor and positive immune nuclei account for roughly 70% of the total, while the two "other" categories together contribute 28 nuclei.
The released code is configured for fewer classes than the dataset defines. Its canonical config lists five type channels — background plus positive tumor, positive immune, negative tumor and negative other — and the repository's parameter audit labels the type decoder "five output channels", against the seven subtypes the paper and the dataset card describe. Whether that is a reduced release configuration or a genuine mismatch between the manuscript and the code is not stated anywhere in the release.
The target use is biomarker quantification in chromogenic immunohistochemistry, where the meaningful number is not a whole-patch positive fraction but positivity resolved per cell compartment: CD47-positive tumor nuclei counted separately from CD47-positive immune nuclei. That distinction is what the subtype vocabulary buys, and it matters for esophageal cancer cohorts profiled for CD47-directed therapy. The dataset and annotation protocol are usable independently of the network, giving IHC groups a labeled starting point where the public alternatives are H&E.
FPNuNet is a worked example of a pattern now common in computational pathology: treat large pretrained vision encoders as fixed feature services and spend the training budget on adapters, fusion and task heads. Holding 98.6% of parameters frozen makes the method cheap to train on the small annotated cohorts IHC work realistically yields. The honest limits are scope and reproducibility. Evaluation is confined to CD47-IHCNuSC, the dataset the same paper introduces, and no result on a cohort the model was not trained on is reported, so transfer to other biomarkers, organs or scanners is untested. The published code release excludes the trained FPNuNet checkpoint, so the manuscript's quantitative results cannot be reproduced from the repository alone.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.