kongnet

KongNet Nuclei Detection Model Architecture [1].

This module defines the KongNet model for nuclei detection and classification in digital pathology. It implements a multi-head encoder decoder architecture with an EfficientNetV2-L encoder. The model is designed to detect and classify nuclei in whole slide images (WSIs).

KongNet achieved 1st on track 1 and 2nd on track 2 during the MONKEY Challenge [2]. KongNet achieved 1st place in the 2025 MIDOG Challenge [3]. KongNet ranked among the top three in the PUMA Challenge [4]. KongNet achieved SOTA detection performance on PanNuke [5] and CoNIC [6] datasets.

Please cite the paper [1], if you use this model.

Pretrained Models:

  • KongNet_MONKEY_1:

    MONKEY Challenge model.

  • KongNet_Det_MIDOG_1:

    MIDOG Challenge lightweight detection model.

  • KongNet_PUMA_T1_3:

    PUMA Challenge model for track 1.

  • KongNet_PUMA_T2_3:

    PUMA Challenge model for track 2.

  • KongNet_CoNIC_1:

    CoNIC model.

  • KongNet_PanNuke_1:

    PanNuke model.

Key Components:

  • TimmEncoderFixed: Encoder module using TIMM models with fixed drop_path_rate handling.

  • SubPixelUpsample: Sub-pixel upsampling module using PixelShuffle.

  • DecoderBlock: U-Net style decoder block with attention mechanisms.

  • KongNetDecoder: U-Net style decoder with multiple decoder blocks.

  • KongNet: Multi-head segmentation model with shared encoder and multiple decoders.

Features:

  • Multi-head architecture for accurate nuclei detection and classification.

  • Efficient inference pipeline for batch processing.

Example

>>> from tiatoolbox.models.engine.nucleus_detector import NucleusDetector
>>> detector = NucleusDetector(model="KongNet_CoNIC_1")
>>> results = detector.run(
...     ["/example_wsi.svs"],
...     masks=None,
...     auto_get_mask=False,
...     patch_mode=False,
...     save_dir=Path("/KongNet_CoNIC/"),
...     output_type="annotationstore",
... )

References

[1] Lv, Jiaqi et al., “KongNet: A Multi-headed Deep Learning Model for Detection and Classification of Nuclei in Histopathology Images.”, 2025, arXiv preprint arXiv:2510.23559., URL: https://arxiv.org/abs/2510.23559

[2] L. Studer, “Structured description of the monkey challenge,” Sept. 2024.

[3] J. Ammeling, M. Aubreville, S. Banerjee, C. A. Bertram, K. Breininger, D. Hirling, P. Horvath, N. Stathonikos, and M. Veta, “Mitosis domain generalization challenge 2025,” Mar. 2025.

[4] M. Schuiveling, H. Liu, D. Eek, G. Breimer, K. Suijkerbuijk, W. Blokx, and M. Veta, “A novel dataset for nuclei and tissue segmentation in melanoma with baseline nuclei segmentation and tissue segmentation benchmarks,” GigaScience, vol. 14, 01 2025.

[5] J. Gamper, N. A. Koohbanani, K. Benes, S. Graham, M. Jahanifar, S. A. Khurram, A. Azam, K. Hewitt, and N. Rajpoot, “Pannuke dataset extension, insights and baselines,” 2020.

[6] S. Graham et al., “Conic challenge: Pushing the frontiers of nuclear detection, segmentation, classification and counting,” Medical Image Analysis, vol. 92, p. 103047, 2024.

Classes

CenterBlock

Center block that applies attention mechanism at the bottleneck.

DecoderBlock

Decoder block with upsampling, skip connection, and attention.

KongNet

KongNet: Multi-head nuclei detection model.

KongNetDecoder

Decoder module for KongNet architecture.

SubPixelUpsample

Sub-pixel upsampling module using PixelShuffle.

TimmEncoderFixed

Fixed version of TIMM encoder that handles drop_path_rate parameter properly.