Slicer Networks

Xiang Chen(陈祥), Hang Zhang, Xiang Chen, Rongguang Wang, Renjiu Hu, Dongdong Liu, Gaolei LiarXiv 2024 · DOI · arXiv

Abstract

Slicer Networks introduce a novel architecture for medical image analysis that leverages the "piece-wise smooth" nature of medical scans. By integrating a learnable bilateral grid into the neural network, the authors achieve a significantly larger effective receptive field (ERF) and an 80%+ reduction in computational demands while outperforming state-of-the-art models in image registration and lesion segmentation.

Introduction

In medical imaging modalities such as MRI, CT, and dermoscopy, objects typically exhibit consistent internal textures separated by sharp, well-defined boundaries—a property known as "piece-wise smoothness." While traditional deep learning backbones like U-Nets are powerful, they face a structural trade-off: deep cascaded decoders are effective at recovering fine details but often shrink the Effective Receptive Field (ERF), limiting the model's global contextual understanding.

Furthermore, high-resolution medical tasks are computationally expensive. Slicer Networks address this gap by explicitly exploiting the low-frequency, edge-preserving characteristics of medical data to improve efficiency and contextual awareness simultaneously.

Methods

The core innovation is the Slicer Network, which replaces or augments traditional decoders with a differentiable bilateral grid process. The workflow consists of three main stages:

  1. Encoder & Reducer: A standard backbone (like ResNet or Vision Transformer) extracts high-level features, which are then compressed into a low-resolution representation.
  2. Learnable Bilateral Grid:
    • Splatting: Features are projected into a higher-dimensional "spatial-range" grid. This uses a guidance map (derived from the raw input image) to ensure that features are grouped by both spatial location and intensity/texture.
  3. Blurring: Learnable convolutional layers operate on this grid. This acts as an edge-preserving low-frequency filter, refining features while maintaining a massive receptive field.
  4. Slicing: The refined grid is "sliced" back into the original image resolution. Because the slicing process is guided by the original high-resolution image, it recovers sharp boundaries that would normally be lost during downsampling.
The detailed Slicer Network architecture, illustrating the flow from feature extraction through splatting, blurring, and slicing operations guided by the raw image.

This approach uses the Deep Directed Accumulator (DeDA) for differentiable splatting and standard Grid Sampling (GS) for slicing, making the entire pipeline end-to-end trainable.

Key Insights and Results

The authors validated Slicer Networks across three distinct medical tasks: unsupervised cardiac Cine-MRI registration (ACDC), lung CT registration, and skin lesion segmentation (ISIC2018).

Effective Receptive Field (ERF)

By slicing directly from a high-dimensional grid rather than passing through multiple upsampling layers, the Slicer Network maintains a much larger ERF. This allows the model to capture broader image context without losing the ability to pinpoint edges.

Visual comparison of the Effective Receptive Field (ERF) across different architectures. The Swin-Slicer (c) demonstrates a significantly larger and more uniform ERF compared to standard decoders.

Performance and Efficiency

The Slicer Network (specifically the "Res-Slicer" and "Swin-Slicer" variants) consistently outperformed established baselines like VoxelMorph, TransMorph, and Swin-Unet:

  • Accuracy: In cardiac registration, Res-Slicer achieved an Average Dice of 79.20%, surpassing the previous best of 77.36%. In skin lesion segmentation, Swin-Slicer reached a 90.02% Dice score, beating the standard Swin-Unet (88.35%).
  • Efficiency: For cardiac registration, the Slicer approach reduced parameter size and Multi-Add operations by 81.0% and 89.0% respectively compared to other bilateral methods, and by over 99% compared to heavy-duty transformer baselines.
  • Zero-Shot Capability: The architecture enabled "zero-shot" learning for lung CT registration, generating accurate deformation fields from keypoints without specific training on those pairs.

Limitations and Future Work

While highly effective for smooth structures, Slicer Networks have limitations:

  1. High-Frequency Detail: As a low-frequency approximator, it is less suitable for tasks requiring the detection of very fine, non-smooth structures like thin curvilinear vessels.
  2. Small Targets: Downsampling in the grid process may cause the model to miss very small objects.
  3. Guidance Map Generation: Future work may involve more sophisticated metric learning for guidance map generation to further improve edge-preserving accuracy.

Overall, Slicer Networks provide a powerful inductive bias for medical imaging, assuming outputs should be piece-wise smooth—a condition met by most segmentation masks and deformation fields in the field.

Summary updated

BibTeX
@misc{chen2024slicer,
  author = {Xiang Chen(陈祥) and Hang Zhang and Xiang Chen and Rongguang Wang and Renjiu Hu and Dongdong Liu and Gaolei Li},
  title = {{Slicer Networks}},
  year = {2024},
  eprint = {2401.09833},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2401.09833},
}
Figure