Loading…
Nuclei detection and classification in histopathology images, using one specialised decoder per cell type over a shared convolutional encoder.
No providers recorded yet. Browse all providers
In a PAS-stained kidney transplant biopsy, a lymphocyte and a monocyte are nearly indistinguishable: PAS leaves the cytoplasm unstained, so the only cue is that lymphocytes tend to be smaller and rounder, and pathologists routinely order an immunohistochemical stain to tell the two apart. A conventional segmentation network asked to label both classes must hold both morphologies in one shared decoder, where the rarer class loses out to the commoner one.
KongNet answers that with a structural change rather than a bigger backbone: a shared encoder feeds a bank of parallel decoders, one per cell class, each seeing only its own class's supervision. Every decoder solves three tasks at once — a map of dilated nuclei centroids, a nuclei segmentation mask, and a nuclei contour mask — so the auxiliary shape tasks push the encoder toward fine-grained morphology while the centroid map remains the object of interest. A global inter-class exclusion term, the product of the class probability maps at each pixel, penalises pixels claimed by more than one class.
The model was developed by the Tissue Image Analytics Centre and VISION Lab at the University of Warwick, with cellular pathology input from University Hospitals of Derby and Burton. It sits alongside CellViT, StarDist and CellNuc-DETR, but takes the opposite bet from the transformer line: a convolutional encoder with task-specific decoders rather than a large pretrained vision backbone.
The encoder is EfficientNetV2-L initialised from ImageNet weights and trained end to end. Each decoder stacks five blocks combining SCSE (spatial and channel squeeze-and-excitation) attention, PixelShuffle upsampling and SiLU activations; a wide variant raises the decoder channel widths to (512, 256, 128, 64, 32). Released checkpoints range from 126M parameters for the detection-only variant, KongNet-Det, to 216M for the ten-class melanoma model, with the headline MONKEY model at 174M. The composite objective combines Jaccard, Dice and focal losses on the centroid map with a BCE-plus-Dice loss on the segmentation and contour maps, weighting the per-class terms either by learned uncertainty or uniformly depending on how skewed the class distribution is.
On the MONKEY Challenge final leaderboard the wide model took first place for overall inflammatory cell detection (FROC 0.3930) and for lymphocytes (0.4624), and second for monocytes (0.2392). It reaches a class-average F1 of 0.653 on CoNIC against 0.616 for StarDist, and 0.674 on PanNuke against 0.618 for CellNuc-DETR and 0.582 for CellViT-SAM-H, whose 699M parameters exceeded the memory of the consumer workstation used for the timing benchmark. The PUMA Challenge checkpoints, fine-tuned from the MONKEY weights, placed third on track 1 (macro F1 0.6466) and second on track 2 (0.2656), with the best lymphocyte score in both. KongNet-Det won the 2025 MIDOG Challenge for mitosis detection with an F1 of 0.7400.
The model suits quantitative pathology workflows that need cell-level counts rather than region-level labels: profiling the tumour microenvironment, scoring tumour-infiltrating lymphocytes, and counting mitotic figures for tumour grading. Its originating use case is transplant pathology, where Banff lesion scoring depends on quantifying inflammatory cells across renal compartments and the lymphocyte–monocyte distinction otherwise requires extra staining. KongNet ships in TIAToolbox's NucleusDetector, which handles tissue masking, tiling and stitching, so a pathology group can run a named checkpoint over a slide without building an inference pipeline.
KongNet is a counterweight to the assumption that nuclei-level pathology tasks are best served by scaling a pretrained vision transformer. Matching or beating foundation-backbone models on PanNuke and CoNIC with an ImageNet-initialised convolutional encoder suggests that decoder structure and loss design still carry weight at this granularity, and a model that runs on a mid-range GPU is usable in laboratories without cluster access. The constraints are real: the Hugging Face card declares every checkpoint CC BY-NC-SA 4.0, so commercial use is excluded even though the inference code is BSD 3-Clause; training code is not published, and each checkpoint is bound to the class taxonomy it was trained on, so a new cell panel requires retraining. The authors also report systematic confusions — apoptotic cells read as lymphocytes, tumour cells as epithelium in epidermal regions — which they attribute to the model's lack of tissue-level context.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.