Computational Data Science Laboratory (), ICAR-CNR

Activities in

SDI+: a Novel Algorithm for Segmenting Dermoscopic Images


SDI+ is an unsupervised algorithm for the segmentation of skin lesions in dermoscopic images, enhancing the capabilities of the SDI algorithm [1]. It is articulated into three steps
  1. extracting preliminary information on possible confounding factors,
  2. accurately segmenting the lesion, and
  3. post-processing the result.
SDI+ results
Performance evaluation of SDI+
Download the SDI+ Matlab code
How to cite the SDI+ software
References

SDI+ results

Experimental SDI+ results have been obtained on the test dataset for the ISIC 2017 lesion segmentation task [2], consisting of 600 dermoscopic images that include skin lesions of different types. Here you can find all the segmentations masks produced by SDI+. Some examples are given below.

A good result (JA=0.967757) Image 16055 SDI+ initial segmentation
Ground truth SDI+ final segmentation

A medium result (JA=0.608539) Image 14219 SDI+ initial segmentation
Ground truth SDI+ final segmentation

One of the worse results (JA=0.205040) Image 15645 SDI+ initial segmentation
Ground truth SDI+ final segmentation

Performance evaluation of SDI+

The evaluation metrics adopted for the ISIC 2017 lesion segmentation task, described in [3], are
Pixel-level accuracy: AC = (TP + TN)/(TP + FP + TN + FN);
Pixel-level sensitivity:SE = TP/(TP + FN);
Pixel-level specificity:SP = TN/(TN + FP);
Dice Coefficient: DI = 2*TP/(2*TP + FN + FP);
Jaccard Index: JA = TP/(TP + FN + FP);
where TP, TN, FP, and FN refer to true positive, true negative, false positive, and false negatives, at the pixel level, respectively. Participants were ranked according to the Jaccard Index.

The complete set of SDI+ performance results is available here, while average performance results achieved by SDI+ and SDI [1] are

AC SE SP DI JA
SDI 0.857 0.692 0.937 0.697 0.592
SDI+ 0.888 0.813 0.927 0.782 0.692

Average performance results achieved by SDI+ on selected subsets of the ISIC 2017 test dataset are

AC SE SP DI JA
Ambiguous (6) 0.648 0.554 0.784 0.465 0.352
Non-Uniform (31) 0.772 0.501 0.936 0.490 0.388
Clear (6) 0.651 0.696 0.672 0.657 0.507
Round (259) 0.839 0.765 0.886 0.727 0.623
Sharp (298) 0.952 0.895 0.970 0.870 0.794

where the ISIC test dataset has been partitioned into five groups defined as follows
  1. Ambiguous (6 images): images whose GT is ambiguous in the shape it delineates;
  2. Non-Uniform (31 images): images that contain either more than one lesion or lesions having non-uniform color, easily perceived as multiple lesions;
  3. Clear (6 images): images that include lesions that are brighter than the surrounding skin;
  4. Round (259 images): images whose GT provides rough lesion contours, having roundness higher than 0.5;
  5. Sharp (298 images): images whose GT provides precise lesion contour, having roundness no higher than 0.5.
The labeling describing the adopted partition is available here.

Download the SDI+ Matlab code

The Matlab code for SDI+ is available here.

How to cite the SDI+ software

If you use the SDI+ software and report results in any publication, we request that you acknowledge this webpage () and the following paper:

M.R. Guarracino, L. Maddalena, , to be published in IEEE Journal of Biomedical and Health Informatics.

References

[1] M. R. Guarracino and L. Maddalena, Segmenting Dermoscopic Images, arXiv:1703.03186, 2017.

[2] N. C. F. Codella, D. Gutman, E. Celebi, B. Helba, M. Marchetti, S. Dusza, A. Kalloo, A. Liopyris, N. Mishra, H. Kittler, and A. Halpern, Skin lesion analysis toward melanoma detection: A challenge at the 2017 International Symposium on Biomedical Imaging (ISBI), hosted by the International Skin Imaging Collaboration (ISIC), arXiv: 1710.05006 [cs.CV], 2017.

[3] D. Gutman, N. Codella, E. Celebi, B. Helba, M. Marchetti, N. Mishra, and A. Halpern, Skin Lesion Analysis toward Melanoma Detection: A Challenge at the International International Symposium on Biomedical Imaging (ISBI) 2016, hosted by the International Skin Imaging Collaboration (ISIC), arXiv:1605.01397 [cs.CV], 2016.