1

I am working with medical images in python. I generate a binary threshold of this image (ImageA, 0, 1 encoded). From this image, I run a detection algorithm that generates another image B (also binary 0,1 encoded) and aims to detect the structures in image A so I can count them.

To better understand what I mean I made a sketch as follows:

Sample image

As you can see, image B failed to recapitulate the structure in the upper right corner. Somehow, I want to test for this. I want to know, how well the algorithm performed that created image B, which aims at contouring the structures it detects in image A.

Is the SSIM algorithm that tests for structural similarity useful for this? Does it take into account to compare the location of structures in the image?

I am working with the structural_similarity() function of skimage.metrics.

Muhammad Mohsin Khan
  • 1,444
  • 7
  • 16
  • 23
Phil
  • 29
  • 5

1 Answers1

0

SSIM algorithm will indeed give you what you need : a good score when img B resembles img A.

However, it is more common to use metrics such as Intersection over Union (see Intersection-over-union between two detections) to measure the performance of an object detection algorithm. Intersection over Union is much more intuitive for binary images than the SSIM that is made to account for local features and textures.

endive1783
  • 827
  • 1
  • 8
  • 18