Model performance

[agdamsbo]

Hi. So I finally had the chance to use this tool on some of our own clinical data.

Comparing to your sample data, I would like to see if I can do anything to improve the model performance, as the lesion masking is quite far from reality. I have ADC, DWI bvalue is 2500. Scanner is a 3T Siemens Vida. The following is the picture report with standard setting. I have run the program on three scans with similar performance.

Do you have any suggestions to increase performance?

[fariaav]

Hi, I think the failure is realted to the contrast of your DWI. It seems that projections tracts have an unusual very high intensity in DWI and low intensity in ADC (which is probably related with your high DWI value of 2,500). The algorithm is interpreting these areas as lesions, in addition to the real infarct. You could try whichever you did not try before: 1) to use your own ADC, instead of us of recalculating it (just add the ADC to your input folder), 2) to recalculate ADC with 1,000 and 2,500 (1,000 is the default, but you have the option to change it). I am not sure if any of these will work because the DWI will still have these high intensities that seem uncommon, but you may try

[agdamsbo]

1) to use your own ADC, instead of us of recalculating it (just add the ADC to your input folder), 2) to recalculate ADC with 1,000 and 2,500 (1,000 is the default, but you have the option to change it).

Hi. Thank you very much for your answer and comments. Our scans are from a clinical setting, where the decision was made to use b 2500 to avoid artefacts on the ADC. In my group we are working on having a dataset of ~500 acute stroke MRIs manually segmented. All on the same scanner with DWI b 2500. Would you be willing to share your code used to train your model? Or be willing to coorporate on fine tuning the model, which I believe would be possible with our dataset? I will be happy to send you an email.