Export only hard lesion masks

Add editor NIfTI download and float32 exports

Preview uploads and export inference runs

Use fixed canvas for mask editor

Keep mask editor layer aligned

Add quick ANTs registration profile

Make ANTs prep timeout and CPU detection robust

I changed prep_utils.py (line 266) so every ANTs subprocess now sets:

Quote full Python 3.10 patch version for HF Space

Restore Python 3.10 pin for Gradio Space

Register + normalize should stay CPU. That ANTs prep step is not the GPU-dependent part,

Enable GPU inference and reset editor on load

Remove python_version to avoid Python 3.1 bug

Use official Python 3.13-slim base image

Fix base image: use Python 3.10

Stabilize mask editor canvas alignment

Auto-prepare active input before segmentation

Make mask editor save slices explicitly

Use supported ImageEditor fullscreen button option

Replace mask editor click hack with ImageEditor

Keep editor case dropdown choices in sync

Keep editor preview visible on invalid click events

Enable experimental mask editor by default

Add slice controls for native and prepared prep viewers

Add feature-flagged experimental mask editor scaffold

Trim skull-edge segmentation artifacts with boundary-aware cleanup

Mask lesion predictions to MNI brain volume

Save inference outputs to mounted persistent storage

Improve soft probability preview with contrast-stretched heatmap

Show thresholded overlay and probability map side by side

Increase Gradio preview image render size

Fix HF model repo downloads for root-level artifacts

Use ARC_ATLAS_v3.1 HF model repo for Space inference

Increase Gradio output viewer sizes for native/registered/inference images

Fix HF dataset input flow and native preview in Gradio

Add HF dataset browser to Gradio inputs and clarify runtime sync caching

Added ZeroGPU-compatible GPU-decorated function in Gradio:

Enable ZeroGPU decorator and switch HF defaults to rmbielski org

Add Gradio app entrypoint and switch Space config from Streamlit

Fix SAM attention softmax rank for CPU inference stability

Auto-preload default ARC_ATLAS V3 model from HF for inference

Add HF dataset model artifact sync for lesion inference

Add Streamlit lesion inference panel and ignore local artifacts

the-brAIn-lab/Approximate_Numeracy added to the dropdown.

the-brAIn-lab/ARC will now appear as a preset option in the "Dataset" dropdown inside the "Browse HF Dataset" tab,

Back to a 50/50 column split

Removed base64, html, and PIL.Image — all three were only needed by the old data URI path. Reload the app and the images should now fill the full container width.

The _render_inline_image function already uses width:100%;height:auto on the <img> tag, so it fills whatever container it's placed in. By switching from a 50/50 column split to a full-width st.container(), both viewers (native-space and registered/normalized) will now render at the full page width. No rebuild needed — just reload the app.

Dockerfile

Implemented a fix for the “runs forever” behavior by removing the two biggest blind spots: repeated runtime sync work and unbounded ANTs subprocess time.