Précision de la segmentation par K-Means

Bonsoir tout le monde. J’aimerais savoir est-ce que c’est possible de mesurer la précision de la segmentation par la méthode du K-means effectué par le plugin Beat/ Cluster Image

Precision with respect to… what? the labels of k-means clustering are arbitrarily assigned.
You need a ground truth or gold standard for that. If you have that kind of data/images, then you could use the Dice index and see for that particular segmentation how well it matched with a known category, but running k-means several times can result in your known category to have a different cluster label.
Alternatively you need to work out, by other means what your kth class might correspond to.

2 Likes

Bonjour, je m’excuses du retard en réalité ce que veux mesurer c’est le taux d’erreur de classification

What I said: use the Dice index:

However you cannot rely on k-means to return (let’s say) consistently 3 classes with the same label every time.
If you have: background, soft tissue and hard tissue with labels 0, 1, 2 and run k-means several times, you might end up with the 3 classes labelled as 2, 1, 0, or 1, 0, 2 or 2, 0, 1 etc, The labels are just labels, not classes. That means that you cannot rely on k-means to detect a class of your choice. Once you have done the clustering, you need to work out what the labels mean, or assign the label correspondences manually. K-means is not a classifier, it is a clustering procedure.

1 Like

D’accord merci bien.