bshepp/vesuvius-nnunet-model
Image Segmentation • Updated
Brian Sheppard PRO
bshepp
AI & ML interests
Reinforcement Learning, Computer VIsion, Fundimental Mathematics, Fundimental AI architecture, AI Safety(for the AI), helping people
Recent Activity
updated
a model 4 days ago
bshepp/vesuvius-nnunet-model posted an
update
6 days ago
Scored 0.390/0.409 on the Vesuvius Challenge Surface Detection — the 1st place team hit 0.614 with nnU-Net. That gap sent us down a rabbit hole.
We built four independent 3D segmentation pipelines to understand what actually matters: standard U-Net, deep supervision + focal loss, multi-scale fusion, and finally nnU-Net itself. Each pipeline taught us something — V1 showed us our GPU was idle 90% of the time (CPU feature bottleneck), V2 proved skeleton recall beats clDice, V3 NaN'd at epoch 55 from FP16 overflow, and nnU-Net is training on HF Jobs right now.
The 1st place post-processing pipeline (binary closing, height-map patching, LUT hole plugging) turned out to be a bigger lever than we expected. We only had access after the competition closed and writeups were published.
Code: github.com/bshepp/volumen published
a model 6 days ago
bshepp/vesuvius-nnunet-model Organizations
updated a model 4 days ago
posted an
update 6 days ago
Post
162
Scored 0.390/0.409 on the Vesuvius Challenge Surface Detection — the 1st place team hit 0.614 with nnU-Net. That gap sent us down a rabbit hole.
We built four independent 3D segmentation pipelines to understand what actually matters: standard U-Net, deep supervision + focal loss, multi-scale fusion, and finally nnU-Net itself. Each pipeline taught us something — V1 showed us our GPU was idle 90% of the time (CPU feature bottleneck), V2 proved skeleton recall beats clDice, V3 NaN'd at epoch 55 from FP16 overflow, and nnU-Net is training on HF Jobs right now.
The 1st place post-processing pipeline (binary closing, height-map patching, LUT hole plugging) turned out to be a bigger lever than we expected. We only had access after the competition closed and writeups were published.
Code: github.com/bshepp/volumen
We built four independent 3D segmentation pipelines to understand what actually matters: standard U-Net, deep supervision + focal loss, multi-scale fusion, and finally nnU-Net itself. Each pipeline taught us something — V1 showed us our GPU was idle 90% of the time (CPU feature bottleneck), V2 proved skeleton recall beats clDice, V3 NaN'd at epoch 55 from FP16 overflow, and nnU-Net is training on HF Jobs right now.
The 1st place post-processing pipeline (binary closing, height-map patching, LUT hole plugging) turned out to be a bigger lever than we expected. We only had access after the competition closed and writeups were published.
Code: github.com/bshepp/volumen
published a model 6 days ago
updated a Space 12 days ago
posted an
update 12 days ago
Post
231
Built an agentic Clinical Decision Support system powered by MedGemma 27B
Paste a patient case → get a full clinical decision support report with differential diagnoses, drug interaction checks (via real OpenFDA/RxNorm APIs), guideline recommendations from a 62-guideline RAG corpus, and automated care gap detection.
6-step agentic pipeline, all streaming in real time:
Parse free-text → structured patient data
Clinical reasoning → ranked differential diagnosis
Drug interaction check → real FDA database queries
Guideline retrieval → RAG over 14 medical specialties
Conflict detection → gaps between guidelines and patient care
Synthesis → comprehensive CDS report
Built with MedGemma 27B, FastAPI, ChromaDB, Next.js. No LangChain — custom orchestrator throughout.
🔗 Try it: https://demo.briansheppard.com
💻 Code: https://github.com/bshepp/clinical-decision-support-agent
🤗 Space: bshepp/cds-agent
Paste a patient case → get a full clinical decision support report with differential diagnoses, drug interaction checks (via real OpenFDA/RxNorm APIs), guideline recommendations from a 62-guideline RAG corpus, and automated care gap detection.
6-step agentic pipeline, all streaming in real time:
Parse free-text → structured patient data
Clinical reasoning → ranked differential diagnosis
Drug interaction check → real FDA database queries
Guideline retrieval → RAG over 14 medical specialties
Conflict detection → gaps between guidelines and patient care
Synthesis → comprehensive CDS report
Built with MedGemma 27B, FastAPI, ChromaDB, Next.js. No LangChain — custom orchestrator throughout.
🔗 Try it: https://demo.briansheppard.com
💻 Code: https://github.com/bshepp/clinical-decision-support-agent
🤗 Space: bshepp/cds-agent
published a Space 20 days ago
liked a dataset about 3 years ago