I am working on developing a segmentation algorithm that can be applied for identifying cancerous regions in CT/MRI-Lung, CT/MRI-brain and CT/MRI-mammography images. I have been trying to evaluate evolutionary algorithm for segmentation that can be applied to all set of images as mentioned above. What I found is while in some cases it gives significantly good results, in others it does not. What is in general a better approach for medical image segmentation :
-> Old School Thresholding approach
-> Machine learning approach (Deep Neural Network, CNN etc)
-> Meta heuristics (Genetic Algorithm, PSO, Simulated Annealing etc)
or any other methods....
Thank you all for your response
@Joachim Pimiskern, this seems to be one interesting approach. Need to explore the aforementioned method specific to my problem.
@Mahdi Javidnasab, We applied K-Means unsupervised algorithm. It clusters the pixels into multiple clusters. However, such method requires a well formed annotated dataset for training. We are still missing out on that.
@Bilal Al-Ahmad , thank you. I have gone through it, there are more recent review papers on the topic, which are also useful.
@Syed Amir Gilani, Thank you.
Hi Ananya,
I think your should define what you want from the image firstly. If you want to simply make the organ's image into several parts corresponding to their gray levels, you can choose a merging/splitting methods, similar with eCognition, or hierarchical grouping. Area morphological segmentation will work well I guess. If interested, we can talk in detail.
Yan
@Yan Li,
Thank you for your response. I am trying to segment out cancerous region from the image. In case of cross sectional lung CT images , first step is to segment out the lung region and then to segment out the cancerous(probable) region. Similar is the case with brain MRI. While the second phase of segmentation works fine once the outer contour of the lung region is segmented. But, it gives undesired result if the cancerous region is large and is attached to the wall of the lungs. In the segmentation process it cancels out actual cancerous region. An example is given as below
- Badges
- Science topic