MIST Stitching

11 min

Requires AMICS 5.2.1 or newer.

MIST (Microscopy Image Stitching Tool) is an advanced field stitching algorithm that produces more accurate mosaic images than the previous SIFT-based stitching method. It uses FFT-based phase correlation to find precise overlaps between adjacent fields, with automatic fallback to stage coordinates when image-based stitching is not possible.

When to Use MIST Stitching

Use MIST Stitching when:

Your measurement has overlapping fields and the default stitching shows visible seams or misalignment

You need sub-pixel accuracy in field alignment for quantitative particle analysis

The auto-adjust stitching produces incorrect results due to repetitive textures or low contrast

MIST Stitching is not needed when:

Fields do not overlap (e.g., sparse point measurements)

The existing auto-adjust stitching already produces satisfactory results

Accessing MIST Stitching

From the Process ribbon Processing tab:

Click the Field Adjuster dropdown in the Tools panel

Select one of the MIST modes:

MIST Stitch — standard mode for most specimens

MIST Low Texture — for specimens with low-contrast or featureless areas

{{FIG: Process ribbon Tools panel showing Field Adjuster dropdown with MIST Stitch and MIST Low Texture options | Process ribbon Field Adjuster dropdown}}

Stitching Modes

MIST Stitch (Standard)

The default MIST mode, suitable for most specimens with moderate to high BSE contrast and visible grain boundaries. It works by:

Computing FFT-based phase correlation between each pair of adjacent tiles

Refining the best overlap candidates using normalized cross-correlation (NCC)

Estimating a global stage model from all pairwise overlaps

Building a maximum spanning tree to position all fields consistently

Best for: Mineral specimens, polished sections, granulated samples with visible grain structure.

MIST Low Texture

An enhanced mode with relaxed matching parameters for specimens that have large areas of uniform grey level (e.g., slag, synthetic materials, low-porosity ceramics).

Low Texture mode differs from standard MIST by:

Considering more candidate peaks during phase correlation (5 vs 2)

Using a lower correlation threshold (0.15 vs 0.50)

Performing more hill-climbing refinement steps (32 vs 16)

Using a narrower high-pass filter to preserve subtle texture features

Best for: Low-contrast specimens, samples with large uniform regions, slag, and ceramic samples.

Note: Even with Low Texture mode, MIST may struggle with completely featureless tiles (e.g., uniform grey background). In these cases, the algorithm falls back to stage coordinates automatically.

Use Secondary Channel (New in 5.3.2)

If your measurement has a secondary detector channel — whatever second channel your SEM software is configured to capture alongside BSE, such as SE, CL, or a second BSE detector — MIST can use it to help stitch samples where the primary BSE channel lacks usable contrast in some areas.

Enable Use Secondary Channel (next to Low Texture Mode in the Field Adjuster dropdown) before running either MIST mode. For each pair of adjacent fields, AMICS correlates both channels independently and keeps whichever one gives the more reliable match for that specific overlap — so a single field can be stitched using BSE in one area and the secondary channel in another, wherever each has more signal.

Best for: Specimens where a companion channel shows structure the primary BSE channel doesn't — for example zircon, where cathodoluminescence (CL) reveals growth zoning that is invisible or flat in BSE.

The toggle is only enabled when the current measurement has a secondary channel available. It roughly doubles stitching compute time when enabled, since both channels are correlated, so it's off by default. Batch processing exposes the same option as a checkbox next to the stitching-mode dropdown in the Batch Process Settings dialog.

Stage Position Stitching (Fallback)

When MIST cannot determine a reliable overlap for a tile pair, it falls back to stage coordinate positioning. This uses the recorded stage positions and the configured field overlap percentage to place fields.

To use pure stage-position stitching:

Click Field Adjuster → Clear Adjustment

When prompted, select Reset to Overlap to position fields using their stage coordinates

Tip: For best stage-position results, enable Cancel Stage Backlash in Investigator Settings before measuring. This ensures that recorded stage coordinates accurately reflect actual positions.

How MIST Works

The algorithm proceeds in four stages:

{{FIG: Diagram showing the 4-stage MIST pipeline: Phase Correlation → Stage Model → Translation Refinement → Global Positioning | MIST pipeline diagram}}

Phase Correlation (PCIAM): For each pair of adjacent fields, compute the FFT cross-power spectrum to find candidate overlap translations

Stage Model Estimation: Use maximum likelihood estimation to determine the typical overlap for the entire grid, filtering out outlier translations

Translation Refinement: Refine each candidate using hill-climbing optimization with normalized cross-correlation scoring

Global Positioning: Build a maximum spanning tree of the best pairwise translations and compute consistent global positions for all fields

GPU Acceleration

If an NVIDIA GPU is available, MIST uses CUDA acceleration for the FFT computation phase, providing approximately 45% faster stitching on large grids. AMICS automatically detects GPU availability and falls back to multithreaded CPU computation if no compatible GPU is found.

After Stitching

After MIST completes:

Field positions are updated in the mosaic view

The classified image reflects the new alignment

Results are saved with the measurement

To revert stitching:

Click Field Adjuster → Clear Adjustment to remove all stitching corrections

Performance

Grid Size	Fields	Approximate Time
4 x 4	16	~2 seconds
8 x 10	80	~3 seconds
42 x 57	1,882	~16 seconds (GPU) / ~29 seconds (CPU)

Comparison with Previous Stitching

Feature	SIFT (Legacy)	MIST
Algorithm	Feature-point matching	FFT phase correlation
Accuracy	Good for high-contrast images	Better sub-pixel accuracy
Low texture	Poor — needs distinct features	Better — Low Texture mode available
GPU acceleration	No	Yes (CUDA)
Fallback	Manual adjustment only	Automatic stage coordinate fallback