Kim et al., 2020, Active learning for accuracy enhancement of semantic segmentation with CNN-corrected label curations: Evaluation on kidney segmentation in abdominal CT

# 세줄 요약 #

1. Recent advances in fully convolutional networks have enabled automatic segmentation, however, high labeling efforts and difficulty in acquiring sufficient and high-quality training data is still a challenge.
2. In this study, a cascaded 3D U-Net with active learning to 3 stages: first, training small dataset with manual labeling ground truth, second, training previous dataset and newly added dataset with CNN-corrected labeling, finally, repeat training all dataset.
3. Active learning was therefore concluded to be capable of reducing labeling efforts through CNN-corrected segmentation and increase training efficiency by iterative learning with limited data. 

 

# 상세 리뷰 #

1. Introduction

• Image segmentation is a fundamental component of medical image analysis.
  • It is a prerequisite for computer aided detection (CAD) and provides information for medical images.
  • First, fully convolutional networks (FCN) have enabled the training of models for semantic segmentation tasks.
  • 3D U-Net has a contracting path and a symmetric expanding path, which has been proven effective for 3D medical image segmentation.
• However, these segmentation methods currently face serious obstacles.
  • Difficulty acquiring sufficient and high-quality training medical image datasets.
  • Variation of human labels.
  • High labeling efforts and costs.
• In this study, we proposed "active learning framework" to reduce labeling efforts as well as increase efficiency with limited training data of medical images.

 

2. Result

• Segmentation results.
  • The average values of the Dice similarity coefficient (DSC) for five subclasses were increased with the completion of each stage. (see table 1)
  • The final segmentation results in the last stage was superior when compared with the nn U-Net using our dataset (see table 2)
    • Segmentation result: manual labeling v.s. CNN-corrected labeling

Table 1. The Dice similarity coefficient (DSC) evaluation. [Kim et al., 2020, SciRep]

Table 2. Comparison of segmentation time between manual and CNN-corrected segmentation. [Kim et al., 2020, SciRep]

• Comparison of time and root-mean-square between manual and CNN-corrected segmentation.
  • The result of the comparison of segmentation time for the five substructures between manual and CNN-corrected segmentation is listed in Table 2.
  • The differences between manual, CNN, and CNN-corrected segmentation by quantitative evaluation in 3D models are presented in Table 3 and Fig 1.
    • The results of CNN-corrected segmentation are observed to highly correspond with those of manual segmentations.

Table 3. Root-mean-square (RMS) evaluation from 3D modeling. [Kim et al., 2020, SciRep]

Fig 1. Result of part comparison analysis in 3D models between (a) manual & CNN segmentation, (b) CNN & CNN-corrected segmentation, (c) manual & CNN-corrected segmentation. [Kim et al., 2020, SciRep]

 

3. Methods 

• Active learning framework
  • Stage 1. we trained the model using a cascaded 3D U-Net with a small amount of training data and the corresponding ground truths were generated by manual labeling at the initial stage.
    • The cascaded architecture was designed to improve segmentation performance using a region proposal network (RPN) prior to segmentation. [Tang et al., 2018, IEEE]
  • Stage 2. the results of the additional data through the trained network were manually corrected labeling.
    • Instead of creating new ground truths from scratch, that means CNN-corrected segmentation.
  • Stage 3. all the data initially used and newly added were used again for subsequent training.

Fig 2. Workflow of active learning framework. [Kim et al., 2020, SciRep]

Fig 3. Data numbers in each stage of active learning. [Kim et al., 2020, SciRep]

Fig 5. 3D U-Net architecture.

 

* Reference: Kim, T., Lee, K., Ham, S. et al. Active learning for accuracy enhancement of semantic segmentation with CNN-corrected label curations: Evaluation on kidney segmentation in abdominal CT. Sci Rep 10, 366 (2020). https://doi.org/10.1038/s41598-019-57242-9