Swell: intrinsic optical signal imaging analysis for spreading depolarization event identification and segmentation

[0.3.2] - 2026-07-21

Analysis workflow

• The global Mask Threshold control now lives with propagation settings. Releasing the slider re-thresholds every prompted frame, while propagated masks generated at a meaningfully different threshold are marked amber and hatched on the canvas and timeline until propagation is rerun.
• Host-launched Analysis windows now restore committed and draft masks, prompts, ground-truth frames, scoped timeline state, and the active frame consistently. Unsaved prompt-only work is included in close warnings.
• Host-to-Analysis mapping signatures reject stale mask, metrics, and project saves after event bounds, baseline scope, or preprocessing settings change. Baseline-scope edits remap saved analysis state; stabilization changes require clearing coordinate-dependent masks, prompts, and ROI geometry.

Image stacks and preprocessing

• Added shared, metadata-aware image decoding for EXIF/TIFF orientation, planar and interleaved color, multi-frame TIFF axes, high-bit-depth grayscale, and consistent Host/Analysis results.
• Stack imports now reject mixed oriented dimensions atomically instead of silently dropping frames, sanitize non-finite pixels throughout preprocessing, and preserve the previous open stack when a replacement fails.
• New stack identity fingerprints prevent a different recording with matching dimensions from being rebound accidentally. When the recorded source folder has changed, an embedded matching stack is preferred when available.
• Baseline subtraction now uses only available pre-event frames and permits a true zero-frame baseline; preview stabilization offsets are converted back to full-resolution coordinates before reuse.

Reliability and quality

• SAM2 cache keys now include source content, visualization content, baseline data, and stabilization offsets, preventing stale prepared frames from being reused across materially different inputs.
• Model runtime initialization validates checkpoint files earlier and fully cleans partial state after failed builds or model switches.
• Improved extreme-aspect-ratio viewport fitting, tooltip scheduling, ROI downsampling, and analysis-window shutdown safety.
• CI now enforces branch coverage and runs image-decoding and persistence tests on macOS and Windows in addition to Linux.

Model/checkpoint compatibility

• No checkpoint format, catalog, or model-selection change in this release; existing managed and local SAM2 selections remain compatible.
• Frame-cache identities changed, so cached prepared frames from earlier versions are not reused when their source or preprocessing content differs.

Platform/backend limitations

• Every frame in a stack must resolve to the same dimensions after EXIF/TIFF orientation is applied. Mixed-size stacks are rejected with the mismatching files listed; resize, crop, or pad them before import.
• CPU fallback behavior is unchanged and remains substantially slower than accelerated SAM2 inference on large stacks.

.swell/migration notes

• No schema-number change in this release. .swell writers still emit schema 3, and schema 2 projects remain loadable.
• Newly saved stack references include optional frame-name and source-content fingerprints. Older projects without these fields remain compatible and gain fingerprints when their stack is rebound or saved through the current workflow.

Known segmentation caveats/regressions

• Mask-generation thresholds are tracked in memory but are not persisted in .swell files. Restored masks therefore begin with an unknown threshold and are treated as fresh until regenerated during the current session.
• Changing Mask Threshold re-thresholds prompted frames on slider release; propagated frames remain visibly stale until propagation is rerun. Ground-truth and manually painted frames are never marked stale.
• No known SAM2 inference or stored-mask regressions in this release.