AGENT SKILL

image-data-utilities

Prepare, validate, convert, and display chest X-ray image inputs while preserving DICOM originals and reporting utility metadata.

License 许可
Apache-2.0
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
image-data-utilities
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Prepare, validate, convert, and display chest X-ray image inputs while preserving DICOM originals and reporting utility metadata.
disable-model-invocation
true

02 资源文件 5

03 SKILL.md 正文

Image-data utilities

Use this skill when a chest X-ray input must be checked, made viewable, or handed between the upload/UI layer and an image-capable tool. It covers the two utility tools:

  • DicomProcessorTool converts a DICOM pixel array to a temporary PNG and returns conversion metadata.
  • ImageVisualizerTool validates a JPG/PNG path and returns the path plus display metadata. Its normal _run path reports a display result; it does not call the matplotlib rendering method in the current implementation.

This skill does not perform model inference. Route interpretation, classification, segmentation, grounding, or report generation to chest-xray-analysis. Route LangGraph construction to agent-orchestration, upload widgets, chat history, and Gradio state to web-interface, and benchmark data preparation to benchmark-evaluation.

Operating contract

  1. Keep the input path and its role explicit: original_path is the preserved source used for archival or a tool that explicitly supports that format; display_path is a viewable PNG/JPG path for the UI or an image-only tool.
  2. Validate a path before invoking a utility. Use the bundled scripts/validate_image_input.py for existence, supported suffix, lightweight file-signature, and optional output-directory checks.
  3. For DICOM, preserve the original file first. Convert a working/display copy with DicomProcessorTool; never replace the original with the generated PNG.
  4. Treat DICOM-derived fields such as PatientID and StudyDate as sensitive. Pass only the metadata needed by the next step, avoid logging it, and do not claim that conversion anonymizes the source.
  5. Inspect the returned tuple for an error key before using image_path. Utility failures are data-preparation failures, not evidence about the patient's image.

See api-reference.md for exact inputs and outputs, data-formats.md for format and privacy rules, workflows.md for path handoffs, and troubleshooting.md for recovery actions.

Choose the path

Existing PNG/JPG/JPEG

  • Validate the file and optional destination directory.
  • Pass the existing image path to ImageVisualizerTool._run when a display record is required.
  • Pass that same path to a downstream model tool only when its contract accepts the format. Keep this utility step separate from model-specific preprocessing.

ImageVisualizerInput documents JPG and PNG. JPEG is commonly represented by .jpg or .jpeg; do not silently accept an arbitrary suffix merely because a reader might open it.

DICOM

  • Validate the original file as DICOM-like input without dumping tags.
  • Call DicomProcessorTool._run(dicom_path, window_center, window_width).
  • If both window parameters are absent, the tool first uses the first value of DICOM WindowCenter/WindowWidth when present. If either is still absent, it uses a min/max scale. If both are present, it clips and scales to uint8.
  • If RescaleSlope and RescaleIntercept are both present, apply them before windowing or min/max scaling (as the tool does).
  • Use returned image_path as the display path and retain the original DICOM path as the source path. The metadata also records original_path and output_path on success.
  • Use the PNG for a model tool that only reads JPG/PNG/PIL/skimage inputs. A DICOM conversion is not a promise that every model tool can read DICOM.

A caller may supply a positive window width and a clinically appropriate center when the source values are known. Do not invent window settings to compensate for a malformed file; fix or reroute the input instead.

Display parameters

ImageVisualizerTool._run accepts image_path, optional title, optional description, figsize=(width, height), and cmap="rgb". For a grayscale CXR, use a grayscale colormap such as "gray" when actual rendering is requested. The implementation treats "rgb" specially and passes other colormap names to matplotlib. If a non-rgb colormap is selected for a multi-channel array, its renderer keeps only the first channel.

The success result is a two-item tuple: output={"image_path": image_path} and metadata containing whether title/description were supplied, the requested figsize and cmap, and analysis_status="completed". A missing file returns an error tuple; do not treat that as a successful visualization.

UI path distinction

The interface stores uploaded data in managed temporary storage. It keeps an original_file_path for the uploaded file and a separate display_file_path. For a .dcm upload it converts only for display; for a regular image the two paths can be the same. UI rendering should use display_file_path, while a message/tool handoff must choose the path according to the receiving tool's format contract. In particular, do not encode raw DICOM bytes as if they were a JPEG data URL. If an image API needs base64, encode the PNG display derivative and use its actual image MIME type, while retaining the original for archival or explicitly DICOM-aware processing.

The utility result from image_visualizer contains the path to show in the UI; it is not a model answer. See web-interface for state and event handling.

Safety and cleanup

  • Do not print full DICOM metadata or embed sensitive tags in titles, descriptions, chat messages, or benchmark records.
  • Use a caller-owned temporary/output directory with appropriate permissions. The DICOM tool creates a unique processed_dicom_<token>.png; it does not anonymize or automatically delete that file.
  • Clean generated derivatives only through an explicit, narrowly scoped cleanup policy owned by the caller. Never delete the original input as part of this skill.
  • Stop on a missing path, unsupported suffix, invalid pixel data, or unwritable destination and report the actionable error. Do not fall back to a random file or silently reinterpret a DICOM as an ordinary image.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-triton-opt

Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-solve

END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-skill

PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-pattern-fusion

Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-orchestrate

Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-feedback

Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).

★ 351