DAGrid: Directed Accumulator Grid

Xiang Chen(陈祥), Hang Zhang, Renjiu Hu, Xiang Chen, Rongguang Wang, Jinwei Zhang, Jiahao LiarXiv 2023 · DOI · arXiv

Abstract

Deep learning models often struggle with medical imaging tasks due to limited data and the complex geometric structures of anatomical features. DAGrid (Directed Accumulator Grid) addresses this by introducing a differentiable module that bakes geometric priors into dense prediction tasks. By shifting from point-wise grid sampling to area-based directed accumulation, DAGrid preserves spatial information more effectively. Our results demonstrate that DAGrid achieves state-of-the-art performance in skin lesion segmentation and cardiac registration while reducing FLOPs by up to 96% compared to transformer-based architectures.

Introduction

Standard convolutional neural networks (CNNs) and Transformers often operate as "black boxes," requiring vast amounts of data to learn the underlying geometry of organs or lesions. While the Directed Accumulator (DA) was previously proposed to incorporate geometric knowledge, its application was restricted to classification or localization. Dense prediction tasks—where every pixel requires a label—remained a challenge. Traditional methods like Grid Sampling (GS) often suffer from information loss because they sample discrete points. DAGrid bridges this gap by providing a geometric-preserving filtering mechanism for pixel-level tasks.

Methodology: The DAGrid Module

The DAGrid architecture operates in three distinct stages to transform and process feature maps geometrically:

  1. Grid Creation: Input feature maps are mapped into a transformed "accumulator space" (e.g., polar or circular coordinates). Unlike standard sampling, DAGrid uses a "Directed Accumulator" approach, ensuring every input pixel contributes to the transformed grid.
  2. Grid Processing: In the transformed space, geometric structures are simplified (e.g., circular boundaries become linear). This allows standard convolutions to act as sophisticated geometric filters, efficiently capturing structural patterns.
  3. Grid Slicing: To return to the original image dimensions, the module "slices" the processed data back. By utilizing homogeneous coordinates for weighted averaging, the module ensures differentiability and prevents information loss during the inverse transformation.
The fundamental difference between Directed Accumulation and Grid Sampling. Note the symmetry between the forward and backward passes of the two methods.

The implementation leverages custom CUDA kernels to maintain high efficiency and support end-to-end training via backpropagation.

Key Insights

Aggregation vs. Sampling

A core innovation of DAGrid is the shift from sampling to aggregation. Traditional grid sampling might overlook critical pixels between sample points. In contrast, DAGrid aggregates all pixels into the transformed space, making the network significantly more robust to fine-scale geometric variations.

Mathematical Symmetry

The authors identify a mathematical symmetry: the gradient flow (backward pass) of a Directed Accumulator mirrors the forward pass of standard Grid Sampling. This relationship is crucial for the module's stability and effectiveness during training.

Long-Range Propagation

Unlike standard CNNs, where information propagation is limited by the receptive field of local kernels, DAGrid enables explicit long-range propagation. Because of the geometric mapping, pixels that are spatially distant in the original image but part of the same geometric structure (e.g., the rim of a heart ventricle) can directly influence one another.

Illustration of long-range accumulation in cardiac MR imaging, showing how information propagates across the image based on geometric priors.

Experimental Results

DAGrid was evaluated on two major medical imaging benchmarks:

  • Skin Lesion Segmentation (ISIC 2018): DAGrid outperformed the transformer-based FAT-Net by 1.0% in Dice score. Remarkably, it achieved this with 70.8% fewer parameters and a 96.8% reduction in FLOPs, highlighting its extreme efficiency.
  • Cardiac Registration (ACDC): In unsupervised cardiac image registration, DAGrid improved the accuracy of left ventricular mass alignment by 8.2% over the VoxelMorph baseline.

Qualitatively, DAGrid produces smoother and more anatomically consistent boundaries, particularly in challenging scenarios with low contrast or artifacts like hair in skin images.

Conclusion

DAGrid demonstrates that explicitly incorporating geometric priors can lead to more efficient and accurate dense prediction models. By replacing discrete sampling with continuous aggregation, it preserves vital spatial information often lost in standard architectures. Future work will focus on extending the module to 3D volumetric data and exploring a wider variety of geometric transforms beyond polar and circular grids.

Summary updated

BibTeX
@misc{chen2023dagrid,
  author = {Xiang Chen(陈祥) and Hang Zhang and Renjiu Hu and Xiang Chen and Rongguang Wang and Jinwei Zhang and Jiahao Li},
  title = {{DAGrid: Directed Accumulator Grid}},
  year = {2023},
  eprint = {2306.02589},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2306.02589},
}
Figure