I have encountered difficulties while attempting to solely utilize the MedNeXt architecture for image segmentation tasks without employing nnUnet's training process. I have observed that the model fails to converge for my specific task, with the Dice coefficient hovering around 0.5. I am seeking insights on the possible reasons behind this issue and would greatly appreciate it if you could share your experiences in this regard. My speculation is that the problem may stem from improper network initialization. Could you kindly advise on how weight initialization should be carried out for MedNeXt? Thank you for your assistance.
I have encountered difficulties while attempting to solely utilize the MedNeXt architecture for image segmentation tasks without employing nnUnet's training process. I have observed that the model fails to converge for my specific task, with the Dice coefficient hovering around 0.5. I am seeking insights on the possible reasons behind this issue and would greatly appreciate it if you could share your experiences in this regard. My speculation is that the problem may stem from improper network initialization. Could you kindly advise on how weight initialization should be carried out for MedNeXt? Thank you for your assistance.