To Whom It May Concern,
I am experiencing a critical technical issue with my CT scan data that is preventing me from proceeding with my thesis analysis.
The primary problem is that the Signal-to-Noise Ratio (SNR) is so low that I cannot perform a reliable thresholding/segmentation process. The bone tissue is indistinguishable from the background noise and artifacts.
Key issues I am facing:
Threshold Inconsistency: It is impossible to select a Hounsfield Unit (HU) or grayscale range that captures the bone structure without including massive amounts of noise.
Loss of Connectivity: When I attempt to isolate the bone, the structural integrity breaks down, resulting in a “fragmented” or “moth-eaten” appearance that does not reflect the actual anatomy.
Boundary Blur: The interface between the cortical bone and surrounding tissue is too blurred for manual or automatic segmentation tools to function correctly.
Since this data is vital for my thesis and I am working against a strict deadline, I urgently need your guidance on:
Which reconstruction filters or denoising algorithms should be applied to the raw data to clean up this noise?
Are there specific artifact reduction settings you recommend to stabilize the image for thresholding?
I have attached a representative slice and dicom showing the severity of the noise. I look forward to your immediate technical feedback.
Best regards,
You could try https://www.openrtk.org/ for reconstruction, I have never used it myself. Maybe also learn if there are some new AI reconstruction methods published
If you have control of the CT scanner you could increase the radiation and the scanning-time to have better images
I doubt you’ll have much success with traditional image processing like thresholding, connectivity, watershed etc.
I think your best bet is an ML approach, such as training an nnU-Net, or one of the existing trained segmenters. I’d give TotalSegmentator to see if it gives you a reasonable starting point. If it seems to work, you could then focus in your specific use-case.
Your problem is not strictly one of noise or artifacts, but of lack of spatial resolution and sensitivity to the different tissues.
Provided that you cannot re-scan you specimen at lower voltage, higher acquisition time, higher resolution and with contrastive agent, you can try ML methods like nninteractive, as written above. Biomedisa can be of help, too.
You can also try to first enhance the dynamics of the scan, then denoise it. I usually use CLAHE for the contrast and Non-local means for the noise. All with Fiji. Non local means is very powerful at removing noise while keeping structures, but requires a long processing time. Try first on a few slices with different parameters.
And if all of that doesn’t work: manual segmentation! (be patient)
