Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LocBAM: Advancing 3D Patch-Based Image Segmentation by Integrating Location Contex

Donnate Hooft, Stefan M. Fischer, Cosmin Bercea, Jan C. Peeken, Julia A. Schnabel

TL;DR

Patch-based 3D medical image segmentation often neglects global patch location, limiting anatomical context. LocBAM introduces a memory-efficient 3D attention mechanism that processes spatial information via three independent 1D gates along the width, height, and depth and fuses them in a U-Net backbone, with Body Part Regression providing normalized location cues and CoordConv as a baseline for comparison. The approach yields consistent Dice improvements across BTCV, AMOS22, and KiTS23, with the largest gains under extremely low patch-to-volume coverage and robustness to location noise. These results demonstrate that incorporating spatial priors via attention can substantially increase reliability and flexibility of patch-based segmentation under memory constraints, enabling smaller patches and more versatile training regimes.

Abstract

Patch-based methods are widely used in 3D medical image segmentation to address memory constraints in processing high-resolution volumetric data. However, these approaches often neglect the patch's location within the global volume, which can limit segmentation performance when anatomical context is important. In this paper, we investigate the role of location context in patch-based 3D segmentation and propose a novel attention mechanism, LocBAM, that explicitly processes spatial information. Experiments on BTCV, AMOS22, and KiTS23 demonstrate that incorporating location context stabilizes training and improves segmentation performance, particularly under low patch-to-volume coverage where global context is missing. Furthermore, LocBAM consistently outperforms classical coordinate encoding via CoordConv. Code is publicly available at https://github.com/compai-lab/2026-ISBI-hooft

LocBAM: Advancing 3D Patch-Based Image Segmentation by Integrating Location Contex

TL;DR

Patch-based 3D medical image segmentation often neglects global patch location, limiting anatomical context. LocBAM introduces a memory-efficient 3D attention mechanism that processes spatial information via three independent 1D gates along the width, height, and depth and fuses them in a U-Net backbone, with Body Part Regression providing normalized location cues and CoordConv as a baseline for comparison. The approach yields consistent Dice improvements across BTCV, AMOS22, and KiTS23, with the largest gains under extremely low patch-to-volume coverage and robustness to location noise. These results demonstrate that incorporating spatial priors via attention can substantially increase reliability and flexibility of patch-based segmentation under memory constraints, enabling smaller patches and more versatile training regimes.

Abstract

Patch-based methods are widely used in 3D medical image segmentation to address memory constraints in processing high-resolution volumetric data. However, these approaches often neglect the patch's location within the global volume, which can limit segmentation performance when anatomical context is important. In this paper, we investigate the role of location context in patch-based 3D segmentation and propose a novel attention mechanism, LocBAM, that explicitly processes spatial information. Experiments on BTCV, AMOS22, and KiTS23 demonstrate that incorporating location context stabilizes training and improves segmentation performance, particularly under low patch-to-volume coverage where global context is missing. Furthermore, LocBAM consistently outperforms classical coordinate encoding via CoordConv. Code is publicly available at https://github.com/compai-lab/2026-ISBI-hooft
Paper Structure (9 sections, 4 figures, 2 tables)

This paper contains 9 sections, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Location-Based Attention Mechanism (LocBAM)
  4. Experiments and Results
  5. Benchmarking
  6. Patch-to-Volume Coverage
  7. Classical Postprocessing vs. Location Context
  8. Robustness to Noisy Location Information
  9. Discussion and Conclusion

Figures (4)

  • Figure 1: LocBAM integrates location context into a (patch-based) segmentation backbone. Overview of LoCBAM integrated into the second block of a 3D U-Net. LocBAM extends HANet’s choi2020cars mechanism to 3D by applying independent 1D attention gates along width, height, and depth, then fusing their outputs via a $1\times1$ convolution.
  • Figure 2: Training curves for BTCV at low-resolution with small patches (Left) and large patches (Right). LocBAM stabilizes performance in low coverage scenarios.
  • Figure 3: Class-wise Dice score improvements over baseline on BTCV. Classes are ordered by size. Dotted lines indicate average gains per method. LocBAM improves performance consistently, especially for large organs.
  • Figure 4: Sensitivity of CoordConv and LocBAM to axial location shifts on BTCV. For small patches, CoordConv degrades sharply while LocBAM remains more stable. Large-patch models are minimally affected.