Hello! I'm a sophomore student. I read your article "h-denseunet" and related codes. Your design is very exquisite, and it's also the only open source and high-precision article in lits liver tumor segmentation. Thank you for your open source. In the process of reading your code, I met some questions and hope to get your answers. Thank you for your reading. The questions are as follows:
In the process of generating training samples, for example, volume-1 generates [3244244], which is cut randomly in your code, so how to calculate the number of randomly generated samples is a problem In this way, it can not be guaranteed that all slices in volume-1 can be randomly arrived and trained, but it is necessary to predict all adjacent slices in the process of test set prediction.
In your code, all (131) are used for training, and there is no verification set, that is, the training process can only be observed through the loss change in the training process, so generally speaking, when to end the training?
During my own experiment, I found that the tumor dice case was much lower than dice global (e.g. dice case-0.65, dice global-0.81). I wondered whether the test set contained CT volume without tumor, so the dice case in this part was equal to 0, that is, it pulled down the tumor dice case
Thank you for reading patiently. I wish you a happy life and happy every day!
Dear author
Hello! I'm a sophomore student. I read your article "h-denseunet" and related codes. Your design is very exquisite, and it's also the only open source and high-precision article in lits liver tumor segmentation. Thank you for your open source. In the process of reading your code, I met some questions and hope to get your answers. Thank you for your reading. The questions are as follows:
In the process of generating training samples, for example, volume-1 generates [3244244], which is cut randomly in your code, so how to calculate the number of randomly generated samples is a problem In this way, it can not be guaranteed that all slices in volume-1 can be randomly arrived and trained, but it is necessary to predict all adjacent slices in the process of test set prediction.
In your code, all (131) are used for training, and there is no verification set, that is, the training process can only be observed through the loss change in the training process, so generally speaking, when to end the training?
During my own experiment, I found that the tumor dice case was much lower than dice global (e.g. dice case-0.65, dice global-0.81). I wondered whether the test set contained CT volume without tumor, so the dice case in this part was equal to 0, that is, it pulled down the tumor dice case
Thank you for reading patiently. I wish you a happy life and happy every day!