Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Weakly Supervised Segmentation of Vertebral Bodies with Iterative Slice-propagation

Shiqi Peng, Bolin Lai, Guangyu Yao, Xiaoyun Zhang, Ya Zhang, Yan-Feng Wang, Hui Zhao

TL;DR

This work tackles vertebral body segmentation from CT scans without expensive voxel-wise labels by introducing a weakly supervised framework that uses four corner landmarks on the mid-sagittal slice. The proposed WISS method combines iterative self-training on the sagittal slice with a slice-propagation strategy to extend segmentation across all slices, aided by confident-prediction selection and dense CRF refinements. Key results show Dice coefficients of $91.7\%$ on mid-sagittal slices and $83.7\%$ on 3D volumes for the lumbar dataset, with improvements over two backbones on a spinal metastases dataset and robustness to noisy weak labels, while substantially reducing labeling costs compared to fully supervised methods. Overall, WISS offers a practical, cost-effective approach to 3D VB segmentation that can generalize to other medical segmentation tasks.

Abstract

Vertebral body (VB) segmentation is an important preliminary step towards medical visual diagnosis for spinal diseases. However, most previous works require pixel/voxel-wise strong supervisions, which is expensive, tedious and time-consuming for experts to annotate. In this paper, we propose a Weakly supervised Iterative Spinal Segmentation (WISS) method leveraging only four corner landmark weak labels on a single sagittal slice to achieve automatic volumetric segmentation from CT images for VBs. WISS first segments VBs on an annotated sagittal slice in an iterative self-training manner. This self-training method alternates between training and refining labels in the training set. Then WISS proceeds to segment the whole VBs slice by slice with a slice-propagation method to obtain volumetric segmentations. We evaluate the performance of WISS on a private spinal metastases CT dataset and the public lumbar CT dataset. On the first dataset, WISS achieves distinct improvements with regard to two different backbones. For the second dataset, WISS achieves dice coefficients of $91.7\%$ and $83.7\%$ for mid-sagittal slices and 3D CT volumes, respectively, saving a lot of labeling costs and only sacrificing a little segmentation performance.

Weakly Supervised Segmentation of Vertebral Bodies with Iterative Slice-propagation

TL;DR

This work tackles vertebral body segmentation from CT scans without expensive voxel-wise labels by introducing a weakly supervised framework that uses four corner landmarks on the mid-sagittal slice. The proposed WISS method combines iterative self-training on the sagittal slice with a slice-propagation strategy to extend segmentation across all slices, aided by confident-prediction selection and dense CRF refinements. Key results show Dice coefficients of on mid-sagittal slices and on 3D volumes for the lumbar dataset, with improvements over two backbones on a spinal metastases dataset and robustness to noisy weak labels, while substantially reducing labeling costs compared to fully supervised methods. Overall, WISS offers a practical, cost-effective approach to 3D VB segmentation that can generalize to other medical segmentation tasks.

Abstract

Vertebral body (VB) segmentation is an important preliminary step towards medical visual diagnosis for spinal diseases. However, most previous works require pixel/voxel-wise strong supervisions, which is expensive, tedious and time-consuming for experts to annotate. In this paper, we propose a Weakly supervised Iterative Spinal Segmentation (WISS) method leveraging only four corner landmark weak labels on a single sagittal slice to achieve automatic volumetric segmentation from CT images for VBs. WISS first segments VBs on an annotated sagittal slice in an iterative self-training manner. This self-training method alternates between training and refining labels in the training set. Then WISS proceeds to segment the whole VBs slice by slice with a slice-propagation method to obtain volumetric segmentations. We evaluate the performance of WISS on a private spinal metastases CT dataset and the public lumbar CT dataset. On the first dataset, WISS achieves distinct improvements with regard to two different backbones. For the second dataset, WISS achieves dice coefficients of and for mid-sagittal slices and 3D CT volumes, respectively, saving a lot of labeling costs and only sacrificing a little segmentation performance.
Paper Structure (6 sections, 2 equations, 3 figures, 2 tables)

This paper contains 6 sections, 2 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Sagittal Slice Segmentation via Self-training
  4. Slice-Propagated Segmentation
  5. Experimental Setup and Results
  6. Conclusion

Figures (3)

  • Figure 1: Overview of the proposed method. We use segmentation output mask incorporating with confident prediction selection and CRF refinements to gradually generate extra training data for VB segmentation. Arrows colored in orange, green and purple represent slice-propagated training at its $1^{st}$, $2^{nd}$ and $3^{rd}$ iterations, respectively. For each iteration, training and refining procedures are repeated in a self-training manner until the model converges. Best viewed in color.
  • Figure 2: Qualitative results for the spinal metastases dataset. For each case, left is the input image and right is the segmentation result. (a) Cervical VBs; (b) Sacral VBs; (c) Thoracic VBs; (d) Lumbar VBs with collapse. Best viewed in color.
  • Figure 3: Qualitative results for the lumbar spine CT dataset. (a) A good case (#4); (b) A bad case (#10). For each case, left is the segmentation result as color overlay with different colors for different instances, and right is the segmentation results as difference maps with oversegmentation errors marked in red and undersegmentation errors in yellow. Best viewed in color.