It is challenging but not impossible. The main idea is to get more confusion matrices, hence points on the ROC curve. If you had scores associated with each image, you could use directly the perfcurve function https://www.mathworks.com/help/stats/perfcurve.html
To understand how it works, check this: http://stackoverflow.com/questions/33523931/matlab-generate-confusion-matrix-from-classifier/33542453#33542453
So the challenge is to assign scores to your 86 images, each of which would tell how close the image is to the true class. Some classifiers return that score, but not K-NN as far as I understand it. Here is one suggestion how you can decide those scores, but you can come up with your own method. http://stackoverflow.com/questions/13642390/knn-classification-in-matlab-confusion-matrix-and-roc?rq=1
