This would be too inefficient (you would have to retrieve the sprite's texture from video memory). Build your "collision map" (a bidimensionnal array of booleans or alpha values, for example) when you load the texture, and then use it for collision detection.