Official implementation for our manuscript "Multicenter privacy-preserving model training for deep learning brain metastases autosegmentation" Published paper
The official Tensorflow version of DeepMedic for our prior Medical Physics paper for brain metastases autosegmentation is publicly available here. This is a Pytorch implementation, which is convenient than the Tensorflow version for further modifications.
The manuscript is officially published in the Green Journal.
Related work: An in-depth technical/experimental survey paper of Incremental Transfer Learning [IEEE Access] [Implementation]:
[1] Huang, Yixing, et al. "An experimental survey of incremental transfer learning for multicenter collaboration." IEEE Access, 2024.
Prior brain metastases autosegmentation paper: [paper] [Implementation]:
[2]Huang, Yixing, et al. "Deep learning for brain metastasis detection and segmentation in longitudinal MRI data." Medical Physics 49.9 (2022): 5773-5786.
This implementation was modified from the source, where we have added our proposed volume-level sensitivity-specificity (VSS) loss, knowledge distillation loss (KDL), naive transfer learning (TL), and learning without forgetting (LWF) training framework. Most instructions from the source still work here.
The main LWF techniques are adpated in the file of train_LWF.py
The overall training and prediction scripts can be found with the names of runMain_GJ.py and runMain_GJ_UKER_Stanford.py. The scripts work for both Windows and Linux environments. But you need to modify the data paths correspondingly.
We have shared several pretrained models, which can be very beneficial for you to evaluate the model performance on your in-house data. You can also apply naive transfer learning (TL) and learning without forgetting (LWF) to improve the model performance. We are glad if you share your evaluation performance or comments/feedback with us.
The pretrained models include UKER independent training, Stanford independent training, TL of UKER -> Stanford, LWF of UKER -> Stanford, mixed training of UKER + Stanford, mixed training of five centers (UKER, Stanford, UCSF, NYU and BraTS). Each setting has 5 repeats with different random seeds.
Free free to open an issue, if you get any problems running this code.
Our environment: Python 3.8, Pytorch 1.12.0+cu113
The Pytorch DeepMedic requires a multicrop library for segment sampling to overcome the class imbalance problem. The original implementation of multicrop from thuyen is outdated. Please find the latest version in my repository, which should be compatible with latest Pytorch versions.
Special thanks to pykao for the original Pytorch DeepMedic implementation. Our implementation was built based on it.