ChangeTracker:Lupus DBP

This page summarizes requirements for longitudinal lesion analysis for Lupus DBP, and resulting discussion on possible modifications required to ChangeTracker to make it possibly a more generic change analysis tool.

Longitudinal lesion analysis for Lupus project

Lupus DBP will face the need of analyzing multi-modal same-subject brain imaging data at 3+ timepoints in order to detect white matter lesions and analyze their progression. At each time point the following data will be available:

• T1
• T2
• FLAIR
• results of lesion classification: (1) labeled connected component filtering of classification/segmentation output + (2) per-label threshold information for each of the connected components

Essentially, pre-thresholded classification output is the per-pixel estimation of likelihood that the pixel belongs to the lesion. Threshold is defined on the per-lesion basis, since its selection depends on the tumor location and other properties.

Requirements to the analysis results are:

• ability to accept multiple lesions, but give control over per-lesion analysis
• ability to go through the available lesion modalities
• CompareView mode for rsults visualization
• naturally, no need for lesion detection functionality
• ability to adjust the classification threshold (per-lesion): should this be done separately for each timepoint?
• distance-based coloring of the change results
• ability to save/load analysis results
• images may be subject to non-rigid deformation due to imaging specifics (coil position) -- non-rigid registration may be required prior to analysis

Consequences for ChangeTracker

Currently, ChangeTracker is essentially designed for change detection in a specific clinical application: meningioma imaging. There may be a need to refactor ChangeTracker to separate out the segmentation functionality from the change detection functionality, to make it more generic.

We can also try to develop two separate workflows within ChangeTracker for Meningioma and Lupus projects, since separating generic change detection functionality may be hard/not possible.

Mock-up workflow for Lupus analysis:

• Step 1: Choose application: Meningioma/Lupus, from here, this is a description of the Lupus-related analysis
• Step 2: Choose number of timepoints, and select with checkbox the types of data available at each timepoint
• Step 3: Select input datasets for each timepoint
• Step 4 (transparent to the user): Register timepoints to the same reference (should this be inside ChangeTracker, or the data has already been registered?), automatically select an ROI for each lesion, so that tumors in all of the timepoints are within the ROI
• Step 5: Create a CompareView layout with a strip for each timepoint, and provide a distance map for each timepoint i that is built in relation to the current binary segmentation at step (i-1). At the same time, in 3d viewer provide color-coded rendering of the growth/shrinking regions in relation to the initial segmentation.

Application-specific modifications suggested for ChangeTracker: provide user controls to

• (1) choose the specific lesion from the labeled set (at timepoint i?);
• (2) expose the threshold slider initialized to the initial setting of the threshold, so that user is able to adjust the threshold and see the compareview updated interactively;
• (3) give ability to save the scene and possibly per-lesion analysis with the specific parameter settings.

Questions

• Are we targeting clinicians or researchers familiar with Slicer?