[lufre1]

Problem

Sometimes a mitochondrion segmentation forms a thin shell around a neighbouring mitochondrion, along the neighbour's outer membrane. In some tomograms the shell even becomes a separate object of its own.

Cause

In _run_segmentation, the watershed heightmap is the normalized inverse distance to the boundaries (boundaries >= boundary_threshold). That distance is 0 on the whole boundary band, so the band is a plateau where the heightmap is exactly 1.0. The order in which the seeded watershed floods voxels of the same height is unspecified. So once a seed's flood reaches a touching mitochondrion's band, it can spread along the outer layers of that band before the band's own mitochondrion claims them.

Fix

Add the boundary probability to the heightmap so that the band is graded rather than flat. The barrier (boundaries > threshold & foreground < threshold) is still set to the maximum of the heightmap. This is a two-line change.

A tie-break by the distance to the nearest seed (as in vesicles.py) was also tested on top of this. It made no measurable difference, so it is left out.

Results

These are 6 tomograms, rerun from cached mitochondria2 predictions with the default parameters. Measurements are at the model resolution (2.87 nm).

• contact: the number of voxel faces between different mitochondria instances.
• max thin: the largest fraction of an instance that a 3-voxel opening removes. Coats are thin, so they show up here.

tomogram	objects before → after	contact before → after	max thin before → after
nav56	6 → 6	0 → 0	0.044 → 0.044
nav60	7 → 7	16 → 12	0.038 → 0.040
nav64	7 → 7	27,926 → 1,089	0.038 → 0.022
nav70	8 → 6 (two objects were 74% and 93% coat)	317,538 → 9,892	0.934 → 0.007
nav91	6 → 6	32,099 → 70	0.190 → 0.043
nav93	9 → 9	57,803 → 8,478	0.127 → 0.127

Instances that don't touch another one are unchanged or nearly so. Segmentations change where mitochondria touch.

Test

TestMitochondriaSegmentation.test_touching_mitochondria in test/test_inference.py runs on synthetic data and needs no network. It uses two balls with overlapping boundary bands. Without the fix, about 19k voxels of one ball (43%) go to the other; with the fix it is under 1%. test_inference.py passes.

🤖 Generated with Claude Code