Today I am presenting DIsoN: Decentralized Isolation Networks for OOD Detection in Medical Imaging at #NeurIPS2025! 🧹💨
📍 Poster #1700, Exhibit Hall C–E
⏰ 4:30–7:30 PM (coming up soon!)
If you’re curious about decentralized OOD detection, come by and say hi! 👋
#AI #DL
🦭 At #NeurIPS2025 in San Diego this week (Dec 1–7), presenting my final PhD project!
If you’re working on medical imaging, foundation models, multimodal learning, federated learning, or OOD detection, let’s meet! Happy to grab a coffee ☕️ or beer 🍺.
DM me to connect! 🌴
🙏 Thanks to my supervisor Prof. @kostaskamnitsas.bsky.social and co-authors @psaha.bsky.social, @harryanthony.bsky.social, Prof. Alison Noble
Excited to present at @neuripsconf.bsky.social - code coming soon!
@ox.ac.uk
@oxengsci.bsky.social
#OOD #ComputerVision #AI #ML #Research
Tested on 12 OOD tasks across 🧴dermatology, 🩻 chest X-ray, ultrasound & 🔬 histopathology.
💥 DIsoN consistently performs strongly against state-of-the-art methods, with higher AUROC and fewer false positives.
Attention bad pun: 🧹 DIsoN cleans up OOD samples like a Dyson 💨
BUT here's the key:
Only model parameters are exchanged between training + deployment, no raw data leaves the training site.
We also add a class-conditional extension (CC-DIsoN):
Compare each test sample only to training samples of its predicted class → stronger OOD performance
DIsoN enables comparing a test sample with the training data distribution, without data transfer!
How?
🔑 We train a binary classifier per test sample to “isolate” it from training data.
📈 The more training steps needed → the more likely the sample is in-distribution.
In medical imaging, safe deployment isn’t just about accuracy.
⚠️ Models must flag unusual scans (artifacts, rare conditions) so clinicians can double-check.
But there’s a problem:
📦 Training data is often private, large, and unavailable after deployment.
Whoop #NeurIPS2025 accepted! 🎉
Meet DIsoN, our 🧹💨 privacy-preserving OOD detector that compares test samples to training data without ever sharing the training data.
We make Out-of-Distribution detection decentralized!
📄Paper: arxiv.org/pdf/2506.09024
🧵👇
The paper will be presented at @wacvconference.bsky.social on March 1 in Arizona🌵
@ox.ac.uk
I am happy that my first post on 🦋 are so exciting news! 🎉
#MedicalImaging #FL #AI #WACV25
Big thank you to my supervisor @kostaskamnitsas.bsky.social and my co-authors: @psaha.bsky.social Wentian Xu Ziyun Liang Daniel Whitehouse Whitehouse David Menon Virginia Newcombe Natalie Voets J Alison Noble
This is the first time we’ve demonstrated that FL can train a single 3D segmentation model for decentralized MRI datasets each with:
🧠 Different brain diseases
📷 Varying MRI modalities
A step forward in training large foundation models for multi-modal MRIs 🙌
🏆 Our results: FedUniBrain was evaluated on 7 MRI datasets with 5 brain diseases.
📊 It achieved promising results across all diseases during training!
Even better, it generalizes to new datasets with unseen modality combinations, something traditional methods fail to do.
We propose the FedUniBrain framework: Train a single model across decentralized MRI datasets with:
✔️ Different brain diseases per dataset
✔️ Different modality combinations per dataset
✔️ No data sharing
Traditional brain segmentation models are disease-specific and rely on predefined MRI modalities for both training and inference. They can’t handle other diseases or scans with different input modalities🚫Plus, patient privacy prevents the creation of big centralized databases🧠
🚀Excited to share our latest work: 🧠FedUniBrain Framework, a necessary step towards training foundation models for multimodal MRIs with Federated Learning, accepted at #WACV25 and selected for an oral!
🔗 arXiv: arxiv.org/pdf/2406.11636
💻 GitHub: github.com/FelixWag/Fed...
🧵1/N