Update download link

[heyufan1995]

No description provided.

[greptile-apps]

Greptile Summary

This PR updates the model download method from wget to the huggingface_hub CLI (hf download) for both NV-Segment-CT and NV-Segment-CTMR, adds the huggingface_hub package (and several brain-preprocessing dependencies) to requirements.txt, bumps torch to 2.1.2 with a matching torchvision in the CTMR package, fixes a bug in run_brain_segmentation.sh where --output_dir was not passed to monai.bundle run, and adds a new ctmr.png visual asset to the root README.

Key changes:

• Download instructions: Both READMEs now use hf download ... --local-dir + mv + rmdir instead of wget, requiring the huggingface_hub package (added to both requirements.txt files).
• Shell script bug fix: --output_dir "$output_dir" is now forwarded to monai.bundle run in run_brain_segmentation.sh, ensuring segmentation output is written to the user-specified directory (consistent with how $preprocess_tmp_seg is resolved in Step 4).
• Dependency updates (CTMR): torch upgraded from 2.0.1 → 2.1.2, torchvision==0.16.2 added, and several brain-preprocessing libraries (pymedio, SimpleITK, pydicom, statsmodels, antspyx) added.
• Documentation: Root README.md table now links class counts to label_dict.json files and embeds a new CTMR architecture image.

Confidence Score: 4/5

• Safe to merge with minor robustness improvements recommended in the download cleanup commands.
• The changes are straightforward: a download-method migration, a clear bug fix in the shell script, dependency additions, and documentation improvements. The only functional concern is that rmdir used after hf download --local-dir could fail if huggingface_hub writes any extra files (metadata, .gitattributes) into the subdirectory — rm -rf would be safer. The "Option 1:" label in the CTMR README without a matching Option 2 is a minor documentation confusion. Everything else (torch upgrade, new deps, bug fix) looks correct.
• NV-Segment-CT/docs/README.md and NV-Segment-CTMR/docs/README.md — both use rmdir which could fail depending on the huggingface_hub version's download behavior.

Important Files Changed

Filename	Overview
README.md	Documentation-only change: adds a hyperlink and "(7 types of tumors)" label to the comparison table, and embeds the new ctmr.png image. No issues found.
NV-Segment-CT/docs/README.md	Replaces wget download with hf download + mv + rmdir pipeline. The rmdir call could fail on versions of huggingface_hub that write extra files (metadata, .gitattributes) into the subdirectory; rm -rf is safer.
NV-Segment-CT/requirements.txt	Adds huggingface_hub (unpinned) to support the new hf download CLI command. Missing newline at end of file.
NV-Segment-CTMR/brain_t1_preprocess/run_brain_segmentation.sh	Bug fix: adds --output_dir "$output_dir" to the monai.bundle run call so segmentation output is written to the user-specified directory, making it consistent with how $preprocess_tmp_seg is resolved in Step 4.
NV-Segment-CTMR/docs/README.md	Replaces wget with hf download pipeline. Same rmdir fragility concern as the CT README; additionally the "Option 1:" label has no corresponding Option 2, which is confusing.
NV-Segment-CTMR/docs/ctmr.png	New binary image asset added; referenced correctly from the root README.
NV-Segment-CTMR/requirements.txt	Updates torch from 2.0.1 to 2.1.2 and adds compatible torchvision==0.16.2; adds brain-preprocessing dependencies (pymedio, SimpleITK, pydicom, statsmodels, antspyx) and huggingface_hub. Missing newline at end of file.

Sequence Diagram

sequenceDiagram
    participant User
    participant Shell as run_brain_segmentation.sh
    participant SynthStrip
    participant Preprocess as preprocess.py
    participant MONAI as monai.bundle run
    participant Revert as revert_preprocess.py

    User->>Shell: --input brain_t1.nii.gz --output_dir results/
    Shell->>SynthStrip: skull strip input → temp/skull_stripped.nii.gz
    SynthStrip-->>Shell: skull_stripped.nii.gz
    Shell->>Preprocess: align to LUMIR template → temp/preprocessed.nii.gz
    Preprocess-->>Shell: preprocessed.nii.gz + meta.json
    Shell->>MONAI: run inference --output_dir results/ (NEW)
    Note over MONAI: Previously output_dir was not passed,<br/>causing segmentation to write to wrong location
    MONAI-->>Shell: results/preprocessed/preprocessed_trans.nii.gz
    Shell->>Revert: revert to original space → results/basename_trans.nii.gz
    Revert-->>Shell: final_output
    Shell-->>User: Segmentation complete

Loading

_{Last reviewed commit: 7dc37ae}

[greptile-apps]

rmdir may fail if extra files are present

rmdir only succeeds on empty directories. Depending on the version of huggingface_hub, the --local-dir download may also create additional metadata files (e.g., .gitattributes, lock files) inside vista3d_pretrained_model/, causing rmdir to fail silently (or halt the script with set -e). Using rm -rf would be more robust:

Suggested change

	rmdir NV-Segment-CT/models/vista3d_pretrained_model
	rm -rf NV-Segment-CT/models/vista3d_pretrained_model

[greptile-apps]

"Option 1" label without a corresponding "Option 2"

The comment labels this step as Option 1: Download using hf and move to expected location, but there is no Option 2 in the instructions. This could leave users confused about whether there is an alternative method. If a second option (e.g., wget or huggingface-cli) was intentionally removed, the Option 1: label should be dropped:

Suggested change

	# Option 1: Download using hf and move to expected location
	hf download nvidia/NV-Segment-CTMR vista3d_pretrained_model/model.pt --local-dir NV-Segment-CTMR/models/ && \
	mv NV-Segment-CTMR/models/vista3d_pretrained_model/model.pt NV-Segment-CTMR/models/model.pt && \
	rmdir NV-Segment-CTMR/models/vista3d_pretrained_model
	# Download using hf and move to expected location
	hf download nvidia/NV-Segment-CTMR vista3d_pretrained_model/model.pt --local-dir NV-Segment-CTMR/models/ && \
	mv NV-Segment-CTMR/models/vista3d_pretrained_model/model.pt NV-Segment-CTMR/models/model.pt && \
	rmdir NV-Segment-CTMR/models/vista3d_pretrained_model

[greptile-apps]

rmdir may fail if extra files are present

Same concern as in NV-Segment-CT/docs/README.md: rmdir is not idempotent and will fail if huggingface_hub's download writes any extra files (metadata, lock files, .gitattributes) into the subdirectory. Prefer rm -rf for reliability:

Suggested change

	rmdir NV-Segment-CTMR/models/vista3d_pretrained_model
	rm -rf NV-Segment-CTMR/models/vista3d_pretrained_model