Extending SEEDS to a Supervoxel Algorithm for Medical Image Analysis

TL;DR

This paper tackles the need for online, fast, and accurate supervoxel generation in medical imaging by extending the SEEDS algorithm to 3D, creating 3D SEEDS. The method preserves the original SEEDS framework while adding 3D-specific boundary handling and a 3D check-splitting function, enabling boundary-safe boundary transfers in sagittal, coronal, and axial directions. Empirical results on BraTS23-Glioma and BTCV show 3D SEEDS outperforms SLIC in speed (about 10×) and over-segmentation quality (up to +6.5% Dice and −0.16% UE), with additional iterations and higher supervoxel counts yielding further gains; 3D SEEDS achieves ADS values approaching those of some deep-learning approaches, illustrating practical utility for online data augmentation. The implementation is open-source (OpenCV-based, C++/Python), enabling easy integration into DL pipelines and offering a flexible speed-versus-accuracy trade-off for medical image analysis.

Abstract

In this work, we extend the SEEDS superpixel algorithm from 2D images to 3D volumes, resulting in 3D SEEDS, a faster, better, and open-source supervoxel algorithm for medical image analysis. We compare 3D SEEDS with the widely used supervoxel algorithm SLIC on 13 segmentation tasks across 10 organs. 3D SEEDS accelerates supervoxel generation by a factor of 10, improves the achievable Dice score by +6.5%, and reduces the under-segmentation error by -0.16%. The code is available at https://github.com/Zch0414/3d_seeds

Figures (7)

• Figure 1: Illustration of the over-segmentation performance. Compared to the SLIC algorithm, 3D SEEDS accelerates supervoxel generation and enhances the achievable Dice score on 13 segmentation tasks across 10 organs.
• Figure 2: Illustration of the 2D check-split function. The deeper blue square represents the portion ($A^{l}_{k}$) of a superpixel ($A_{k}$) to be transferred to another superpixel ($A_{n}$) in the horizontal direction (from left to right). The lighter blue square indicates the remaining part of $A_{k}$, while the orange square belongs to any other superpixel except for $A_{k}$. The 2D check-split function returns false if cases (a), (b), or (c) occur; otherwise, it returns true.
• Figure 3: Illustration of $N_{\mathcal{A}^{l}_{k}}$ in the boundary term. (a) The deeper blue square represents a single pixel $A^{l}_{k}$ in the superpxiel $A_{k}$, which is to be transferred to another superpixel in the horizontal direction (from left to right). $N_{\mathcal{A}^{l}_{k}}$ is computed within the $3 \times 4$ area surrounding $A^{l}_{k}$, ignoring the orange square. (b) $N_{\mathcal{A}^{l}_{k}}$ is computed in the 3D space.
• Figure 4: Illustration of the 3D check-split function. The deeper blue cube represents the portion ($A^{l}_{k}$) of a supervoxel ($A_{k}$) to be transferred to another supervoxel ($A_{n}$) in the sagittal direction (from left to right). The lighter blue cube represents the remaining part of $A_{k}$, while the orange cube belongs to any other supervoxel except for $A_{k}$. The 3D check-split function returns false if cases (a)-(p) occur; otherwise, it returns true.
• Figure 5: Illustration of MRI Modalities v.s. Brain Tumor Subregion Over-Segmentation Performance. The evaluation uses the default 3D SEEDS configuration with 1000 supervoxels.