GP-based Image Segmentation (GPIS) with Applications to Biomedical Image Segmentation (c) Louis Charbonneau and Nawwaf Kharma, 2009.

GP-based Image Segmentation (GPIS)

withApplications to Biomedical

Image Segmentation

(c) Louis Charbonneau and Nawwaf Kharma, 2009

Problem statement


• How do we select an optimal sequence of low-level image operators (& parameters) to get the segmented image?

Segmentation example: cell nuclei




…




Model description• We use Cartesian GP:– Primitive operators are clearly defined, their right combination is the problem

– CGP allows for an easy interpretation of the resulting sequence

– Segmentation is a class of problems without one perfect solution; CGP can handle this


System objectives • Effectiveness: segmentation should be correct• Efficiency: The smallest number of operations• Transparency: operation sequences should be easy to understand


System objectives (cont.)• Segmentation should be doable without a priori information (except for training ground truths)• Generality: effective on wide classes of images• Ease of Use: Minimal human intervention



Fitness criterion


Fitness criterion


Fitness criterion


Crossover


Mutations (I)


Mutations (II)



Data


1026 images, 512 x 384 pixels

120 images, 340 x 780 pixels

System settings, database 1


Pixel segmentation accuracy, database 1


Cell segmentation accuracy, database 1


Statistical results, database 1


Example of evolved program, database 1




System settings, database 2


Pixel segmentation accuracy, database 2


Cell segmentation accuracy, database 2


Statistical results, database 2




Intermediate steps of evolved program, database 2






Superimposed input + evolved program


GPIS on other types of images


Lane detection tree detection

GPIS on other types of images


Intra-cellular content of Wright-stained white blood cell images

Conclusion• CGP was able to adapt to the complexity of input images:– A short program was evolved to solve the easy problem – a longer program was evolved to solve the harder problem

• Operator pool can be extended with specialized operators• Injection was a reliable means of maintaining population diversity


Conclusion• A training window approach is very effective for operator refinement

• A small but accurate set of ground truths is enough to evolve segmentation algorithms without a priori information on the images


GP-based Image Segmentation (GPIS) with Applications to Biomedical Image Segmentation (c) Louis Charbonneau and Nawwaf Kharma, 2009.

Documents

nawwaf kharma

thisc louis charbonneau

understandc louis charbonneau

crossoverc louis charbonneau

datac louis charbonneau

pixelsc louis charbonneau

mutations ic louis charbonneau

cell nucleic louis charbonneau