Blob detection returns scikit-image sigma as if it were a blob radius

[PierreRaybaut]

Summary

sigima/tools/image/detection.py::__blobs_to_coords returns scikit-image's sigma as if it were a blob radius:

def __blobs_to_coords(blobs: np.ndarray) -> np.ndarray:
    cy, cx, radii = blobs.T
    coords = np.vstack([cx, cy, radii]).T
    return coords

The third column produced by skimage.feature.blob_log, blob_dog and blob_doh is the scale σ at which the blob was detected, not its radius:

• for blob_log and blob_dog, the blob radius is approximately √2 · σ;
• for blob_doh, the blob radius is approximately σ.

The conversion factor therefore differs per method, and none is currently applied. The docstrings of find_blobs_log / find_blobs_dog / find_blobs_doh add to the confusion by describing the min_sigma / max_sigma parameters as "the minimum/maximum blob radius in pixels".

Why this is tracked separately

Applying the correct conversion changes user-visible results: the radius of the circles drawn over detected blobs, and the size of the ROIs auto-generated around them via create_image_roi_around_points. It is therefore a behaviour change, not a documentation fix.

The documentation-only part is handled in the "assorted minor issues" issue.

Scope

• Decide on the convention: return a true radius (per-method conversion) or keep sigma and rename the column accordingly
• Apply the per-method conversion factor in __blobs_to_coords or in each find_blobs_* wrapper
• Re-baseline sigima/tests/image/blobs*_unit_test.py and any ROI-size assertions
• Check downstream consumers in DataLab and DataLab-Web (blob overlays, generated ROIs)
• Release note announcing the behaviour change

[PierreRaybaut]

The documentation-only part of this problem is covered by #50.