Less is More: Sample Selection and Label Conditioning Improve Skin - - PowerPoint PPT Presentation

▶
less is more sample selection and label conditioning
SMART_READER_LITE
LIVE PREVIEW

Less is More: Sample Selection and Label Conditioning Improve Skin - - PowerPoint PPT Presentation

Sep 09, 2023 512 likes •768 views

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation Vincius Ribeiro, Sandra Avila, Eduardo Valle RECOD Lab., DCA, FEEC, University of Campinas (UNICAMP), Brazil RECOD Lab., IC, University of

slide-1
SLIDE 1

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation

ISIC Workshop @ CVPR 2020

Vinícius Ribeiro¹, Sandra Avila², Eduardo Valle¹

¹RECOD Lab., DCA, FEEC, University of Campinas (UNICAMP), Brazil ²RECOD Lab., IC, University of Campinas (UNICAMP), Brazil

slide-2
SLIDE 2

Handling the Inter-Annotator Agreement for Automated Skin Lesion Segmentation

Available at arXiv e-prints

2 Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

The Impact of Annotation Quality on Deep Learning for Skin Lesion Segmentation

Available at Unicamp's Scientific and Intellectual Production Repository Links at https://vribeiro1.github.io/publications/

slide-3
SLIDE 3

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

Sample Selection & Inter-annotator Agreement

slide-4
SLIDE 4

4

Lesion Attributes

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

slide-5
SLIDE 5

Different levels of granularity for image annotation 1. Flood fill controlled by the annotator 2. Manual polygon tracing 3. Fully automated validated by human

5

Different Levels of Annotation

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

slide-6
SLIDE 6

6

Different Opinions by Physicians

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

slide-7
SLIDE 7

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

Data Elimination with Label Conditioning

slide-8
SLIDE 8

8

Sample Conditioning

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

Original ground truth Opening Opening + Convex Hull

slide-9
SLIDE 9

9

Sample Selection

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

slide-10
SLIDE 10

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

Materials & Methods

slide-11
SLIDE 11

ISIC Archive 13,779 images with segmentation 2,233 images with multiple segmentations

11

Two train datasets: All samples - 2,233 images Best samples - 1,808 images (κ ≥ 0.5) Three test datasets: ISIC Archive - 2,000 images with single segmentation PH2 - 200 images Dermofit - 1,300 images

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

Dataset

slide-12
SLIDE 12

ISIC Archive 13,779 images with segmentation 2,233 images with multiple segmentations

12

Two train datasets: All samples - 2,233 images Best samples - 1,808 images (κ ≥ 0.5) Three test datasets: ISIC Archive - 2,000 images with single segmentation PH2 - 200 images Dermofit - 1,300 images

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

Dataset

slide-13
SLIDE 13

ISIC Archive 13,779 images with segmentation 2,233 images with multiple segmentations

13

Two train datasets: All samples - 2,233 images Best samples - 1,808 images (κ ≥ 0.5) Three test datasets: ISIC Archive - 2,000 images with single segmentation PH2 - 200 images Dermofit - 1,300 images

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

Dataset

slide-14
SLIDE 14

14

Training

Random Selection DCNN Label Conditioning Jaccard Index Result

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

slide-15
SLIDE 15

15

Validation

DCNN Label Conditioning Jaccard Index Result 1 Result 2 Best Result

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

slide-16
SLIDE 16

16

Testing

DCNN Label Conditioning Jaccard Index Result

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

slide-17
SLIDE 17

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

Results

slide-18
SLIDE 18

18 Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

slide-19
SLIDE 19

19 Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

slide-20
SLIDE 20

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

Conclusion

slide-21
SLIDE 21
  • Segmentation ground-truths for skin lesion imaging present

substantial inter-annotator disagreement.

  • Withholding samples with the largest disagreement may result in

significantly improved performance.

  • Removing details from the segmentation masks may improve the

results.

  • In the future, we hope to extend our findings to samples with single

ground-truth masks, increasing its applicability.

21

Conclusion

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

slide-22
SLIDE 22

All the skin lesion images and segmentation masks from this presentation were extracted from the ISIC Archive dataset, available at https://isic-archive.com/

22 Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020

slide-23
SLIDE 23

Thank you!

Vinícius Ribeiro vinicius.ribeiro1@gmail.com Sandra Avila @sandraavilabr Eduardo Valle dovalle@dca.fee.unicamp.br

Code & Data:

Less is More: Sample Selection and Label Conditioning Improve Skin Lesion Segmentation, Ribeiro et al., ISIC Workshop @ CVPR 2020