search

MIC-DKFZ / nnDetection

nnDetection is a self-configuring framework for 3D (volumetric) medical object detection which can be applied to new data sets without manual intervention. It includes guides for 12 data sets that were used to develop and evaluate the performance of the proposed method.
Apache License 2.0
527 stars 88 forks source link

[Question] Train on 2D medical images #36

Open zoliseress opened 2 years ago

zoliseress commented 2 years ago

:question: Question

Hi! Is it worth to try experimenting with the nnDetection framework on 2D medical images, or is it designed (as the description says) solely for "3D (volumetric) medical object detection"?

mibaumgartner commented 2 years ago

Hi @zoliseress ,

thanks for your interest in nnDetection. The current version does not implement a 2D pipeline because all of our data sets in the training/validation pool were 3D. Since the 2D networks produced worse results on most of our 3D data sets we did not run additional experiments in that direction.

In some preliminary experiments (on pure 2D data sets), we also observed that some fixed parameters do not generalize well between 3D and 2D due to different Anchor Distributions and potentially different Network requirements (usually 2D images have a much higher resolution and thus networks with a bigger field of view could be preferable there).

Best, Michael

zoliseress commented 2 years ago

Thank you for the detailed answer, Michael!

I was excited about your network's performance on our data, but sadly it seems that they are incompatible. Anyway, keep up the good work!

Zoli

andreped commented 2 years ago

2D support would be of interest to me as well.

I have previously experimented with RetinaNets, YOLO networks, and Faster R-CNN.

Testing your design is the next step for me, to properly benchmark our designs which we have tuned ourselves.

Found this rather late... I had spent quite some time getting this to install on colab with pip (working now though). Should've checked issues earlier...

Kunlei-Hong commented 2 years ago

how to initialize the model？ I didn't find any tips

ArpanGyawali commented 2 weeks ago

2D support would be of interest to me as well.

I have previously experimented with RetinaNets, YOLO networks, and Faster R-CNN.

Testing your design is the next step for me, to properly benchmark our designs which we have tuned ourselves.

Found this rather late... I had spent quite some time getting this to install on colab with pip (working now though). Should've checked issues earlier... Hey, did you install this wihour using docker? can you help me?