LKCell: Efficient Cell Nuclei Instance Segmentation with Large Convolution Kernels

TL;DR

The paper tackles the problem of accurate and efficient cell nuclei instance segmentation in histology images. It proposes LKCell, a U-Net–style architecture that uses large convolution kernels carried by a pretrained LK-encoder and a single LK-Decoder with multi-task heads to capture multi-scale context while maintaining low parameter counts. On PanNuke, LKCell achieves an $mPQ$ of $0.5080$ and a $bPQ$ of $0.6847$ at only $21.6\%$ of the FLOPs of the previous leading method, demonstrating strong efficiency and accuracy. The method generalizes to MoNuSeg, offering a practical, compute-efficient solution for clinical deployment with robust segmentation performance across tissue types.

Abstract

The segmentation of cell nuclei in tissue images stained with the blood dye hematoxylin and eosin (H$\&$E) is essential for various clinical applications and analyses. Due to the complex characteristics of cellular morphology, a large receptive field is considered crucial for generating high-quality segmentation. However, previous methods face challenges in achieving a balance between the receptive field and computational burden. To address this issue, we propose LKCell, a high-accuracy and efficient cell segmentation method. Its core insight lies in unleashing the potential of large convolution kernels to achieve computationally efficient large receptive fields. Specifically, (1) We transfer pre-trained large convolution kernel models to the medical domain for the first time, demonstrating their effectiveness in cell segmentation. (2) We analyze the redundancy of previous methods and design a new segmentation decoder based on large convolution kernels. It achieves higher performance while significantly reducing the number of parameters. We evaluate our method on the most challenging benchmark and achieve state-of-the-art results (0.5080 mPQ) in cell nuclei instance segmentation with only 21.6% FLOPs compared with the previous leading method. Our source code and models are available at https://github.com/hustvl/LKCell.

Figures (4)

• Figure 1: (a) Receptive Field. By appropriately enlarging the size of the convolutional kernel, the network can effectively capture the overall structure of the cells without introducing excessive computational load. (b) Performance of LKCell. We illustrate the computational efficiency and performance metrics of LKCell compared to previous methods. LKCell achieves state-of-the-art performance with minimal FLOPs.
• Figure 2: Comparison with previous best methods. (a) represents a typical Hover-Net graham2019hover shaped model consisting of three decoder branches, each producing three different maps. On the other hand, (b) represents our model, which consists of a single decoder and three separate segmentation heads for different outputs. This significantly reduces the parameter and computational complexity of the model.
• Figure 3: Architeture of LKCell. We present the overall architecture of LKCell. The encoder is composed of a pre-trained model ding2023unireplknet with large convolution kernels and is connected to the decoder through skip connections. The decoder consists of four LKCellBlocks. Each LKCellBlock is a combination of Large Kernel and Small Kernel, along with components such as BatchNorm, GELU, ReLU, and 1$\times$1 convolution. Postprocessing technique is employed to match nuclei types and refine nuclei segments.
• Figure 4: Comparison of Segmentation Results.We compare the segmentation results of 19 different types of cell nuclei using LKCell on the PanNuke dataset with Ground Truth and obtain highly accurate instance segmentation results.