Sensitivity/specificity is defined for a two class problem, so to calculate these quantities for a multi-class problem you need to re-formulate, for each class extracting a table consisting of two classes, class X and not class X, the calculation then follows by the book.  Overall accuracy I would calculate (naively) as the sum of the true positive diagonal divided by the total number of cases in the confusion matrix. But this is a little slippery in the literature, for example see:

http://www.alanfielding.co.uk/multivar/accuracy.htm or

http://hal.archives-ouvertes.fr/docs/00/61/13/19/PDF/17_ICIT11_VL.pdf

Sensitivity/specificity is defined for a two class problem, so to calculate these quantities for a multi-class problem you need to re-formulate, for each class extracting a table consisting of two classes, class X and not class X, the calculation then follows by the book.  Overall accuracy I would calculate (naively) as the sum of the true positive diagonal divided by the total number of cases in the confusion matrix. But this is a little slippery in the literature, for example see:

http://www.alanfielding.co.uk/multivar/accuracy.htm or

http://hal.archives-ouvertes.fr/docs/00/61/13/19/PDF/17_ICIT11_VL.pdf

In matlab you can use the "plotconfusion" function.

AKbar's answer is correct. However for generalization you may look into the following links

http://www.cs.odu.edu/~mukka/cs495s13/Lecturenotes/Chapter5/recallprecision.pdf

http://stats.stackexchange.com/questions/51296/how-to-calculate-precision-and-recall-for-multiclass-classification-using-confus

For a multi-class problem , the only thing you can do is a average ROC curve , by using one vs all approach , but as akbar suggested confusion matrix is a better way to analyse such problems , as it gives a fair idea regarding class to class misclassification , which might be useful to improve the algorithm.

how to calculate sensitivity and specifisty for three class