implementation of MIDL2024 submission on implicit neural obfuscation by Mattias P. Heinrich (University of Lübeck) and Lasse Hansen (EchoScout.ai) Paper page on OpenReview
Re-identification is a severe risk to patients' privacy even if proper anonymisation of images is undertaken. Our proposed method advances the state-of-the-art in effective medical image obfuscation strategies with regards to the following three main points:
- robust generative model, by adapting recent work neural implicit representation and compression for video sequences to the obfuscation of a subset of a X-ray collection.
- novel strategy for k-anonymity that only moderately affects visual image quality while substantially reducing re-identification risks
- alleviates the strong requirements of prior work that are based on simultaneous availability of multiple scans per patients at each data provider
The key concept of our proposed implicit neural obfuscation strategy is as follows: A subset of input chest X-rays serve as target for a neural reconstruction decoder that comprises learnable instance embeddings (D-dimensional vector for each data point) and convolutional weights. The reconstructions are supervised with a loss based on SSIM. During inference a k-anonymity mixing is introduced that aims to obfuscate patient information by adding latent code information from other patients.
This repository implements the first two parts (NeRV fitting and Semantic segmentation with obfuscated images). It is in part based on the exellent work of NeRV by Haochen and uses the SSIM implentation of Evan Su.
To install clone the repo and create + activate a new virtualenv virtualenv env; source env/bin/activate
Next get the requirements by runnning pip install monai zipfile36 scikit-image tqdm imageio wget torchvision
Now you can either register for, and download the Kaggle data used in the paper RANZR CLiP and preprocess it accordingly (detailed instructions will be provided at a later point) OR you can use a public demo dataset Montgomery County CXR using the following command:
python create_demo_xray.py
Next you can fit the NeRV in three chunks of 48 images and train a SegResNet (CUDA is required):
python fit_nerv_ssim.py img_label_demo.pth 0 46
python train_segresnet.py img_label_demo.pth 0 46 8 122 1
The implicit neural obfuscation is very quick, the segmentation network may take 15-20 minutes. Finally you can evaluate the outcome on 16 validation cases using:
python run_inference_to_png.py
This produces a validation Dice of approx. 81% and a folder with 16 example outputs (left ground truth image and label, right NeRVed scan and predicted label).
For the third part: re-identification risk assessment, we provide another demo-dataset that is derived from the NLST2023 challenge dataset of Learn2Reg, which comprises 220 patients with two lung CTs, which we convert into digitally reconstructed radiographs (DRRs). The provided jupyter notebook gives details on how we train the siamese network with InfoNCE and evaluate re-ID risk using 15-fold test-time augmentation. In line with our more ellaborate results in the paper, we achive a substantial improvement in reducing such risks with hires-image baseline: top1 53.12% and top5 72.92% being reduced to NeRV (rho=0.08) with top1 39.58% and top5 53.12%.
If you have any questions feel free to email me at firstname.heinrich@uni-luebeck.de ToDo: provide code for pre-processing RANZR-CLiP data.