google/embeddinggemma-2
Feature Extraction • 0.7B • Updated • 29.2k • 1.29k
🤝 Open to Collab
Md Sazidur Rahman
yokko123
·
AI & ML interests
Medical Imaging, Ischemic Stroke, Segmentation
Recent Activity
posted an update about 3 hours ago
🧠 Our IEEE BHI 2026 paper is on arXiv, and the model is on the Hub.
Bi-temporal Image-driven Acute Stroke Evolution Analysis
Admission CT perfusion tells you what brain tissue looks like now. Follow-up DWI tells you what happened to it. We register the two and intersect their labels into six outcome-aware ROI classes, then ask one question: does admission CTP already carry the tissue's eventual fate?
For penumbra, yes. Salvaged and infarcted penumbra separate consistently in feature space (Δ̃cos = 0.146, p < 0.05), and it holds across first-order statistics, GLCM radiomics and CNN embeddings alike. For core, it barely separates by subsequent fate, which fits tissue already irreversibly injured on arrival.
The largest separation wasn't inside the hypoperfused area at all. Tissue that looked normal at admission but later infarcted sits far from healthy contralateral brain (Δ̃cos = 0.460). The same pattern replicates on ISLES'24, across different centres, scanners and annotation pipelines.
Released today:
🤗 nnU-Net core/penumbra segmentation model, 40-channel 4D CTP in, 5-fold ensemble
Dice 0.71 penumbra / 0.30 core on a held-out test set
https://huggingface.co/yokko123/ctp-core-penumbra-nnunet
📄 https://huggingface.co/papers/2602.07535
💻 https://github.com/yokko123/bi-temporal-ctp-dwi-code
🔗 https://yokko123.github.io/bi-temporal-ctp-dwi/ updated a model about 3 hours ago
yokko123/ctp-core-penumbra-nnunet liked a model 2 days ago
google/embeddinggemma-2Organizations
None yet
