The Datasets tab (DatasetsPage.qml, route Datasets) manages immutable dataset versions scoped to the currently open project. Each dataset captures curated image patches, semantic segmentation masks, class definitions, and WSI-level split metadata so training runs remain reproducible and free of data leakage.
Datasets view showing immutable dataset versions, status badges, WSI/class/tile counts, split ratios, sizes, and Create dataset button.
Availability & Prerequisites
Active Project Required: Yes (Open a project to browse its dataset versions).
Source Data: At least one whole-slide image with ground-truth annotations (drawn in Viewer or imported via QuPath) is required to build a non-empty segmentation dataset.
Step-by-Step User Workflows
1. Create a New Dataset Version
Click Create dataset in the page header to launch DatasetWizardDialog.
Walk through the guided wizard steps:
Specify dataset name and semantic class selection/mapping.
Select source WSIs and verify annotations coverage.
Configure patch/tiling parameters (tile dimensions, target magnification/MPP, stride/overlap, background/tissue filtering) and Train / Validation / Test split ratios (enforced at the WSI level so patches from one slide never leak across splits).
Launch dataset generation and inspect validation diagnostics. When finished, the dataset transitions to Ready with a sequential version (v1, v2, …).
DatasetWizardDialog step 1 showing the WSI selection table, format, and imported annotation classes.
2. Preview or Delete a Dataset Version
Preview: Use the row View action to open DatasetViewerDialog (datasetPreviewDialog) to inspect generated statistics, class balance, sample patches, and storage artifact metadata.
Delete: Click the row Delete action to open Delete dataset version? (datasetDeleteDialog), which permanently removes the dataset record and generated patch artifacts after confirmation.
flowchart LR
WSI["Annotated WSIs in Active Project"] --> Wizard["DatasetWizardDialog"]
Wizard --> Split["WSI-Level Split (Train / Val / Test)"]
Split --> Extract["Patch + Alignment Mask Extraction"]
Extract --> Verify["Verify Class Semantics & Artifacts"]
Verify --> Ready["Immutable Dataset Version (Ready)"]
Ready --> Train["Selectable in Training Tab"]
UI Controls & Table Columns Reference
Header & Empty State
Element
Label / Object Name
Condition
Behavior
Page Title
Datasets
Always visible
Description updates to "Immutable dataset versions for <projectName>" when a project is active
Create dataset
createDatasetButton (Create dataset)
Disabled if !hasActiveProject
Calls datasetWizard.startNewDataset() and opens DatasetWizardDialog
No-project placeholder
datasetsNoProjectState
Visible when !hasActiveProject
Prompts user to open a project in Projects
Datasets Table (datasetsTable)
Column Key
Label
Meaning
name
Name
Dataset version human-readable name (e.g., "Tumor glands — baseline")
Preview (DatasetViewerDialog) and Delete (datasetDeleteDialog)
Modal Dialogs
Dialog
Component
Purpose
Dataset Creation Wizard
DatasetWizardDialog (datasetWizardDialogInstance)
Multi-step slide selection, class verification, tiling parameters, WSI-level split assignment, and build progress
Dataset Preview
DatasetViewerDialog (datasetPreviewDialog)
Inspect dataset composition, class distribution, and artifact path details
Delete Dataset Confirmation
datasetDeleteDialog
Confirms irreversible deletion of the dataset version and its generated artifacts
Technical Reference: ML & Data Integrity Rules
Zero Train/Val/Test Leakage: Dataset splits are partitioned strictly by whole-slide image rather than individual tile patches so neighboring tiles from one patient slide never appear in both training and validation/test sets.
Deterministic Validation & Test: Validation and test splits exclude stochastic augmentations and preserve exact image-to-mask spatial alignment and class index ordering.
Immutability: Once materialized, a dataset version (v1, v2, …) is treated as immutable input provenance for downstream Training jobs and Models registry records.