Skip to main content

Inference Explorer

The Inference Explorer tab offers a detailed visual and tabular exploration of all predictions made during a validation run. It is primarily used to assess individual segmentation outcomes and compare them against available ground truth annotations for each study.

Model and Dataset Cards​

At the top of the page, two summary cards provide metadata on the model and dataset involved in the selected validation:

  • Model Card: Displays details such as model name, version, modality, anatomy, class mapping, framework (e.g., PyTorch), and deployment status. The date of the last validation run is also shown.
  • Dataset Card: Lists the dataset name, source institution, image modality, format (e.g., DICOM), annotation ID, dataset size, and number of studies included in the validation run.

Prediction Table​

Below the summary cards, a table lists all studies processed in the validation:

  • Columns include:
    • Transaction ID
    • Status (e.g., Completed)
    • Study ID
    • Modality
    • SOP Classes
    • Instance / Series count
    • Tags and Source Tags

Each row represents a unique prediction result tied to a specific study.

Study Viewer​

Clicking a row expands a dual-panel viewer:

  • Left Panel: Displays the predicted segmentation result in color overlay. Each anatomical structure is listed with toggleable visibility.
  • Right Panel: Shows the original medical image (e.g., MRI or CT scan) without overlay.

Additional options include:

  • Open in the Playground: Launches the study in the Playground module for advanced interaction.
  • Select Annotation: Allows you to load ground truth annotations, enabling side-by-side comparison with model predictions.

This interface is especially useful for manual error inspection, class-wise quality checks, and visual validation.

Key Use Cases​

  • Review segmentation results per study
  • Compare model output with ground truth
  • Navigate predictions within a validation batch
  • Identify and isolate underperforming predictions visually

✅ Tip: Use this page to spot edge cases or labeling inconsistencies by visually comparing automatic segmentations against the ground truth.