PACTHub← Back to home

Guide to integrating the existing AI algorithm

The project team's AI algorithm is already complete. This guide explains how to connect the existing work to the dataset website and assemble the reproducibility information needed for public release. It does not require redesigning or reimplementing the algorithm.

1. Separate presentation from availability

The website should describe each of the following separately:

  • The task the algorithm addresses, the data it supports, and its known limitations.
  • The completion status of the algorithm itself.
  • Whether code, weights, papers, and documentation are public, and whether access must be requested.
  • Website integration status. “Algorithm complete” does not automatically mean “the current website can run the algorithm.”

The current framework has no online inference backend and includes no algorithm source code, model weights, performance results, or user-data upload feature. Resources that have not been supplied must remain pending integration. Empty links must not be labeled “Download” or “Run online.”

2. Assemble the minimum public materials

Extract the following from the existing project rather than creating new technical claims:

  1. The algorithm name and a short introduction, specifying the input signals or volumes and the outputs.
  2. The actual code repository, release tag or exact commit, and corresponding paper or technical description.
  3. Model weights that may legally be distributed, with their configuration, version, file size, and checksum.
  4. Environment requirements: operating system, language/framework, dependency lockfile, CPU/GPU requirements, and GPU memory needs.
  5. Verified installation and inference steps, including input directory structure, configuration, and output paths.
  6. Code and weight licenses, citation instructions, intended scope, and known failure cases.

You can continue improving the local framework while public-release approval is pending. Do not publish files first and seek approval afterward.

3. Map to the completed data standard

Algorithm documentation should reference the existing data standard and its version, and verify:

  • Array axis order, coordinate system, data types, units, and voxel spacing.
  • Which wavelengths, sampling rates, array geometries, and acquisition or reconstruction parameters the algorithm actually requires.
  • How fields in the data standard map to the website index. The website index is not the model's input-file format.
  • The actual resampling, normalization, cropping, masking, and missing-value procedures.
  • Output shape, units, spatial alignment, and postprocessing.

If the algorithm does not need a field, explicitly mark it as not applicable. Do not invent parameters to satisfy a page template. If the website template lacks necessary fields, extend and validate the mapping layer rather than silently changing the existing standard.

4. Verify reproducibility

Run the full workflow on a real example approved for reproducibility, and record the actual results:

  • Versions and checksums for the data, code, weights, and configuration.
  • Environment, hardware, random seeds, and any nondeterministic operations.
  • The actual commands and order for preprocessing → inference/reconstruction → postprocessing.
  • Expected output names, formats, dimensions, units, and acceptable error tolerances.
  • For evaluation metrics: definitions, reference ground truth, splits, aggregation methods, and execution conditions.
  • For speed, memory, or GPU measurements: the measurement method and hardware conditions.

The DEMO records in this repository are synthetic metadata for page testing. They cannot establish algorithm performance or produce a valid scientific benchmark. Results not yet supplied should be labeled “Pending provision/verification.” Do not insert fabricated PSNR, SSIM, accuracy, or runtime values.

5. Integrate with the website

  1. Update algorithm.name, algorithm.status, algorithm.repository_url, and algorithm.weights_url in data/site-config.json. Keep other unknown fields as null. Verify repository and weight links in practice.
  2. Update the algorithm section using text, figures, and results approved by the project team. Continue identifying illustrative figures as illustrations.
  3. Distinguish “Read documentation,” “View code,” “Get weights,” and “Request access” with clear labels.
  4. Update data versions, algorithm versions, and citations together. Do not associate the latest data with incompatible older weights by default.
  5. Check every link and its access conditions in a fresh browser session, then complete the release checklist.

The current configuration provides the algorithm name, summary, status, repository, and weights entry points. Papers, environment details, commands, metrics, and version compatibility can initially be maintained in this guide's Markdown and HTML versions. If you add structured configuration fields, update the page's rendering logic too; adding JSON alone does not make it appear automatically. The algorithm resource package in downloads is configured separately. Its url, version, size_label, and sha256 must describe an actual released package.

This framework can link to verified external code, weights, and documentation. Online inference would require a separately designed backend, access control, computing resources, privacy handling, and service limits. The current static website does not automatically provide these capabilities.

6. Integration information still needed from the project team

This is a checklist of materials needed for website integration, not an assessment of research progress.