Correction Learning for Medical Image Segmentation

Guang Zhang, Shuhao Dong, Hui Xu, Hongyu Zhang, Yongjian Wu, Yanwei Zhang, Xiaoming Xi, Yilong YinIEEE Access 2019 · OpenReview

Abstract

Breast tumor segmentation is a critical component in the diagnosis of breast cancer. However, ultrasound imaging presents significant challenges, including intensity inhomogeneity and shadowing artifacts. This paper proposes a robust segmentation method based on "correction learning." Unlike traditional approaches that strive for complex, perfect initial models, correction learning identifies and automatically fixes erroneous segmentations using a two-stage process: coarse segmentation followed by specialized correction modules. The framework includes Internal Tumor Block Correction (ITBC) to address intensity issues and Boundary Block Correction (BBC) to mitigate shadowing artifacts. Experimental results on a multi-device database demonstrate the method's superior accuracy and robustness.

Introduction

In medical imaging, achieving perfect segmentation on the first pass is often hindered by the inherent noise and artifacts of clinical data. Researchers typically respond by increasing model complexity, which can lead to overfitting and poor generalization. This study introduces a paradigm shift: "Correction Learning." The core philosophy is that detecting and fixing errors is often more feasible than preventing them entirely. By employing a "good enough" initial model and a targeted correction system, we can achieve high-fidelity results even in challenging ultrasound environments.

The Challenges of Ultrasound Segmentation

Automated breast ultrasound analysis faces two primary hurdles:

  1. Intensity Inhomogeneity: Tumors often contain internal regions with intensities similar to healthy tissue. This leads to "holes" or false negatives within the segmented tumor area.
  2. Shadowing Artifacts: Dense tumors can cast dark shadows beneath them. These shadows mimic the appearance of tumor tissue, causing the segmentation to "leak" into the background, resulting in false positives.
Examples of breast tumor segmentation in ultrasound images, showing (a) a benign lesion and (b) a malignant lesion with irregular boundaries.

Proposed Methodology: Correction Learning

The proposed framework operates in two distinct phases.

1. Coarse Segmentation

The process begins with an Active Contour Model (ACM). By incorporating a bias field assumption and local intensity clustering properties, the ACM generates an initial, rough outline of the tumor. This stage provides the foundation for subsequent refinement.

2. Correction Learning Modules

The novelty of this approach lies in its two correction components:

  • Internal Tumor Block Correction (ITBC): To fix errors from intensity inhomogeneity, the image is partitioned into superpixels using Simple Linear Iterative Clustering (SLIC). A Support Vector Machine (SVM) is then trained to identify missed tumor blocks within the initial contour, effectively filling the "holes."
  • Boundary Block Correction (BBC): To address shadowing, BBC utilizes spatial logic. Since shadows typically appear below the lesion, the module identifies and prunes dark blocks that lack the neighborhood characteristics of true tumor tissue, preventing the segmentation from leaking into shadow regions.

Experimental Evaluation

The method was validated using a self-constructed database of 186 cases (135 benign, 51 malignant) collected from four different ultrasound manufacturers (ALOKA, Toshiba, GE, and SIEMENS). This diversity ensures the algorithm is robust across different hardware platforms.

The Correction Learning (CL) framework demonstrated significant performance gains over traditional and recent machine learning approaches (such as MLF1 and MLF2):

  • Accuracy: Achieved 0.923 for benign and 0.912 for malignant lesions.
  • Jaccard Index: Reached 0.4906 for malignant tumors, substantially outperforming competing methods which often hovered between 0.12 and 0.28.
  • Cross-Device Robustness: While other methods saw performance declines when moving between different ultrasound machines, the CL method remained stable, proving its clinical viability.

Conclusion

The "Correction Learning" approach offers a robust alternative to hyper-complex segmentation models. By focusing on error detection and correction through superpixel analysis and spatial logic, the framework effectively handles the most common artifacts in breast ultrasound. Future work will explore the integration of deep learning architectures to further enhance the correction phase.

Reference: Zhang, G., et al. "Correction Learning for Medical Image Segmentation." IEEE Access, vol. 7, 2019.

Summary updated

Abstract

Breast tumor segmentation is useful to diagnose breast cancer. However, challenges, such as intensity inhomogeneity and shadowing artifacts arise in this task. To address these two issues, this paper proposes a robust ultrasound image segmentation method based on correction learning. At first, a novel idea of correction learning is introduced. In contrast to traditional methods that develop the complex models to obtain accurate segmentation results, correction learning aims to detect the erroneous segmentation in advance and correct this automatically by only using simple method. The proposed correction learning method mainly involves two steps: coarse segmentation and correction learning. First, an active contour model is firstly constructed for coarse segmentation by introducing assumption of bias field and local intensity clustering property. Then, correction learning is developed to address the erroneous segmentation and to improve the segmentation performance. In this paper, correction learning mainly contains Internal Tumor Block Correction (ITBC) and Boundary Block Correction(BBC). In order to correct the erroneous segmentation caused by intensity inhomogeneity, the internal tumor blocks detection model based on Simple Linear Iterative Clustering(SLIC) and Support Vector Machine (SVM) is learned to detect these segmented blocks. Based on the detection result, the incorrect segmented blocks can be corrected. In addition, BBC is proposed to correct the erroneous segmentation of boundary blocks which are caused by shadowing artifacts. Our experiment results on the constructed database demonstrate the effectiveness and robustness of the proposed method.
BibTeX
@inproceedings{zhang2019correction,
  author = {Guang Zhang and Shuhao Dong and Hui Xu and Hongyu Zhang and Yongjian Wu and Yanwei Zhang and Xiaoming Xi and Yilong Yin},
  title = {{Correction Learning for Medical Image Segmentation}},
  booktitle = {IEEE Access},
  year = {2019},
  url = {https://openreview.net/forum?id=torm69g3C4},
}
Figure