Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DLA-Count: Dynamic Label Assignment Network for Dense Cell Distribution Counting

Yuqing Yan, Yirui Wu

TL;DR

DLA-Count tackles the challenge of counting densely packed, morphologically variable cells by introducing three core innovations: K-adjacent Hungarian Matching (KHM) for density-aware label assignment, Multi-scale Deformable Gaussian Convolution (MDGC) for adaptive morphology-aware feature extraction, and Gaussian-enhanced Feature Decoder (GFD) for effective multi-scale fusion. The model uses a VGG16-bn backbone to generate multi-scale features, then regresses a cell-position map from which local maxima yield cell centers, with KHM providing adaptive, Gaussian-weighted matching within density-dependent radii. Extensive experiments on four benchmarks (ADI, MBM, VGG, DCC) demonstrate state-of-the-art performance in dense scenarios, with substantial improvements in MAE and MSE, and ablation studies confirm the critical contributions of both KHM and MDGC. The approach offers a scalable, end-to-end solution for automated, robust cell counting in diverse biomedical images, potentially accelerating pathology and drug discovery workflows by reducing manual counting effort and enhancing reproducibility.

Abstract

Cell counting remains a fundamental yet challenging task in medical and biological research due to the diverse morphology of cells, their dense distribution, and variations in image quality. We present DLA-Count, a breakthrough approach to cell counting that introduces three key innovations: (1) K-adjacent Hungarian Matching (KHM), which dramatically improves cell matching in dense regions, (2) Multi-scale Deformable Gaussian Convolution (MDGC), which adapts to varying cell morphologies, and (3) Gaussian-enhanced Feature Decoder (GFD) for efficient multi-scale feature fusion. Our extensive experiments on four challenging cell counting datasets (ADI, MBM, VGG, and DCC) demonstrate that our method outperforms previous methods across diverse datasets, with improvements in Mean Absolute Error of up to 46.7\% on ADI and 42.5\% on MBM datasets. Our code is available at https://anonymous.4open.science/r/DLA-Count.

DLA-Count: Dynamic Label Assignment Network for Dense Cell Distribution Counting

TL;DR

DLA-Count tackles the challenge of counting densely packed, morphologically variable cells by introducing three core innovations: K-adjacent Hungarian Matching (KHM) for density-aware label assignment, Multi-scale Deformable Gaussian Convolution (MDGC) for adaptive morphology-aware feature extraction, and Gaussian-enhanced Feature Decoder (GFD) for effective multi-scale fusion. The model uses a VGG16-bn backbone to generate multi-scale features, then regresses a cell-position map from which local maxima yield cell centers, with KHM providing adaptive, Gaussian-weighted matching within density-dependent radii. Extensive experiments on four benchmarks (ADI, MBM, VGG, DCC) demonstrate state-of-the-art performance in dense scenarios, with substantial improvements in MAE and MSE, and ablation studies confirm the critical contributions of both KHM and MDGC. The approach offers a scalable, end-to-end solution for automated, robust cell counting in diverse biomedical images, potentially accelerating pathology and drug discovery workflows by reducing manual counting effort and enhancing reproducibility.

Abstract

Cell counting remains a fundamental yet challenging task in medical and biological research due to the diverse morphology of cells, their dense distribution, and variations in image quality. We present DLA-Count, a breakthrough approach to cell counting that introduces three key innovations: (1) K-adjacent Hungarian Matching (KHM), which dramatically improves cell matching in dense regions, (2) Multi-scale Deformable Gaussian Convolution (MDGC), which adapts to varying cell morphologies, and (3) Gaussian-enhanced Feature Decoder (GFD) for efficient multi-scale feature fusion. Our extensive experiments on four challenging cell counting datasets (ADI, MBM, VGG, and DCC) demonstrate that our method outperforms previous methods across diverse datasets, with improvements in Mean Absolute Error of up to 46.7\% on ADI and 42.5\% on MBM datasets. Our code is available at https://anonymous.4open.science/r/DLA-Count.

Paper Structure

This paper contains 14 sections, 4 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Problem Definition and Method Overview
  4. Problem Definition
  5. Method Overview
  6. K-adjacent Hungarian Matching (KHM) Algorithm
  7. Multi-scale Deformable Gaussian Convolution (MDGC)
  8. Experiments
  9. Experimental Setups
  10. Datasets.
  11. Implementation Details.
  12. Comparison with State-of-the-Art Approaches
  13. Ablation Study
  14. Conclusion

Figures (4)

  • Figure 1: Overall architecture diagram of DLA-Count.
  • Figure 2: Core design of the K-adjacent Hungarian Matching (KHM) Algorithm.
  • Figure 3: The architecture of our Multi-scale Deformable Gaussian Convolution (MDGC)-(a), internal modules-(b)(c)(d). (a) shows the overall MDGC structure with parallel processing paths; (b) illustrates the Gaussian Bottleneck with multi-scale AGConv operations; (c) details the FusionAttention mechanism for adaptive feature recalibration; and (d) visualizes the receptive field pattern generated by the Adaptive Gaussian Convolution (AGConv).
  • Figure 4: Visualization across multiple datasets.