What about ROC?
To segment an image into healthy and non-healthy regions, how do I check which method best segmented the image into these categories? Is there any tool in MATLAB for it, or some other way to find the solution?
Note: the segmentation is based on multilevel threshold.
"Standard approaches for evaluating segmentation results given ground-truth segmentation include the Hausdorff distance, Dice coefficient [9], and the Jaccard index [16]" [ http://www.cs.sfu.ca/~hamarneh/ecopy/miccai2008b.pdf ]
Hi Sheema,
There are many methods for validating the segmentation results, but in all cases the selection of a method depends on the source image.
if a natural image is used in the segmentation then the best way is to use manual delineation of objects. This is applies to other type of images such as MRI images which require expert opinion. However, for multicomponent images this is more difficult. The method of verification varies between different segmentation methods e.g. healthy segmented object is an object with a threshold > than a value, otherwise it is unhealthy.
I suggest reading some of my papers to check how the segmentation results re verified depending on image type and characteristics. Good luck
You should have ground truth images done by some expert or yourself and calculate e.g. the true positive/all pixel count and the true negative pixels/all pixel count for your images.
"Standard approaches for evaluating segmentation results given ground-truth segmentation include the Hausdorff distance, Dice coefficient [9], and the Jaccard index [16]" [ http://www.cs.sfu.ca/~hamarneh/ecopy/miccai2008b.pdf ]
as mentioned above, dice factor is a common evaluation metric. tanimoto metric is another common metric. but PSNR and error image are used in some segmentation papers as far as I know.but you should have a manually segmentation image ,prepared by an expert,as the groundtruth , first.
Dear Sheema,
The best way is to use classification, but you need to have a large sample of images where the non-diseased and the diseased segmented areas may be compared. Another way is to extract texture or multiscale AM-FM features from the segmented areas (healthy and diseased) and find significant differences between features. We have done quite a lot of work towards these directions. See our pubs below. There is also a software tool available to download with which you can segment, perform texture analysis and extract features from the segmented areas.
1. C.P. Loizou, V. Murray, M.S. Pattichis, M. Pantziaris, A.N. Nicolaides, C.S. Pattichis, “Despeckle filtering for multiscale amplitude-modulation frequency-modulation (AM-FM) texture analysis of ultrasound images of the intima-media complex”, Int. J. Biomed. Imag., vol. 2014, Art. ID. 518414, 13 pages, 2014. (open access)
2. C.P. Loizou, C. Theofanous, M. Pantziaris, T. Kasparis, “Despeckle filtering software toolbox for ultrasound imaging of the common carotid artery”, Comput. Meth. & Progr. Biomed., vol. 114, no. 1, pp. 109-124, 2014. Best paper, editor’s choice
3. C.P. Loizou, C.S. Pattichis, M. Pantziaris, T. Tyllis, A. Nicolaides, “Quality evaluation of ultrasound imaging in the carotid artery based on normalization and speckle reduction filtering,” Med. Biol. Eng. Comput.,” vol. 44, no. 5, pp. 414-426, 2006.
4. C.P. Loizou, C.S. Pattichis, C.I. Christodoulou, R.S.H. Istepanian, M. Pantziaris, A. Nicolaides “Comparative evaluation of despeckle filtering in ultrasound imaging of the carotid artery,” IEEE Trans. Ultras. Ferroel. Freq. Contr., vol. 52, no. 10, pp. 1653-1669, 2005
For qantitaive analysis of your segmentation result you can use a parameter named Dice Similarity Coefficient (DSC). For this, first get the ground truth image of your test image; and then compute true positive, true negative, false positive and false negative parameter using your output image and ground truth image. Finally compute the Dice Similarity Coefficient (DSC) using below equation.
DSC = (2*TP)/((FP+TP)*(TP+FN))
where T-True,P-postiev,F-False,N-Negative.
For some standard image database like Berkeley image database, ground truth are given. For other test you can generate ground truth images using software like "imageJ".
if you are using experts, dont forget to include analysis of intra and inter user variability to actually make the best interpretation of your results. There are simpler metrics such as overlap, euclideana nd hausdorff distances that you can use , or a combination of metrics, but in all cases inter and intra variability needs to be studied.
There are various methods to quantify the segmentation results. Comparing with ground truth image is one of them. The classification method to choose is depends on the type of image and a classification accuracy required. Not necessary to use computationally complex method all the time.
- Badges
- Science topic