How to classify brain tumor.

Hi, i am working on a project Brian tumor Segmentation and classification, i tried but didn't find a training Dataset for malignant and benign, please help me how to classify and which classifier is best for tumor classification, also if possible please provide me the link to download the Dataset for malignant and benign tumor. Thanks

[Manu BN]

Here is a code that can help: https://in.mathworks.com/matlabcentral/fileexchange/55107-brain-mri-tumor-detection-and-classification

[sowmya budiga]

Will this code run on version R2020a

[Walter Roberson]

No, the code uses svmtrain() which was replaced with fitcsvm(). The last release it worked in was R2017b.

Answer 1

[Walter Roberson]

http://www2.imm.dtu.dk/projects/BRATS2012/data.html

And read http://uk.mathworks.com/matlabcentral/answers/29574-automatic-detection-of-melanoma#answer_37922

[Muhammad Shoaib]

i want Dicom image and these image are in other format

[Walter Roberson]

http://www.osirix-viewer.com/datasets/ has two

http://sites.duke.edu/dblab/dce-mri-test-images/ has a lot (!); a summary is at https://idash.ucsd.edu/rider-neuro-mri

[sam CP]

Is it possible to detect the tumor area by using the classifier??

[Walter Roberson]

That is the goal, yes, that by training on a number of known results that predictions could be made about images whose results were unknown. People have had varying success with doing this.

Our experience was that results were notably more accurate if you worked with MRS (Magnetic Resonance Spectroscopy) rather than MRI (Magnetic Resonance Imaging). Accuracy for automatic classification on images tends to hover a bit below what an experienced person would say; at most match what an experienced person would say. To go beyond that, to see patterns that an experienced observer would not pick up, we found we had to use MRS. And we found that if you did MRS then imaging was not always necessary: we had notable successes for cancer detection by urine, for examine.

[sam CP]

I have classified the tumor (Benign or Malignant ) by using the classifier. For the classification purpose , i have used the set of known result( database of Benign and Malignant tumor). I have these database (Known result of MRI with brain tumor, and the features are extracted from those images)But how will be a SVM classifier works for detecting the tumor area?

[Walter Roberson]

The arrangement you describe will not tell you the tumor area.

However, you could divide your image into blocks and classify each block separately, and then approximate the tumor area as the sum of the areas of the blocks classified tumors.

Dividing into blocks might lead to problems if you are using shape descriptors as a block that is inside the tumor might not have the right shape to be detected.

[sam CP]

I have used K-means clustering algorithm for segmenting the tumor contained MRI(It divided the image into blocks or clusters). what is the syntax or method for detecting the tumor area by using a classifier ?(KNN classifier)

[Walter Roberson]

If clusterindex is the name of the variable that tells you which cluster was associated with which pixel, and TumorIndex is the cluster number that was used for the tumor, then

sum(clusterindex == TumorIndex) * AreaOfOnePixel

[sam CP]

'imIDX' is my k-means clustered image. Here imIDX image contain 4 blocks(4 colours). One block contain tumor. Then i have extracted 12 features from the imIDX image. My question is that how can i detect the tumor section from the imIDX image by using KNN classifier?

[Walter Roberson]

"My question is that how can i detect the tumor section from the imIDX image by using KNN classifier?"

What does the medical literature say are the physical characteristics of that kind of tumors? For example if medical dissection has shown that inevitably that kind of tumor grows into a blob that spells out "THIS IS A TUMOR" in Hangul (Korean), then you have something to go on.

[sam CP]

Physical characteristics of tumor are Asymmetry ,Shape deformation, Intensity etc

[Walter Roberson]

Okay, so one partial step would be use regionprops to measure the Eccentricity and MeanIntensity of each region, discarding those with low eccentricity (under the assumption that "Asymmetry" implies "not round") or whose intensity does not match whatever it is that is required.

... and then you could choose randomly among all of the ones that remain and call that the tumor.