Brain volume change is the mean Jacobian determinant over the baseline brain mask. Its boundary check repeats the mean with the mask eroded 6 mm: -1.98 % full against -3.05 % eroded over the interval (disagreement 0.49 points per year: green ≤ 0.5, amber ≤ 1.5, red beyond). A change confined to the outer rim is rim-dominated: a mask difference between the two visits, or cortical change the eroded mask leaves out; the calibration cannot separate the two because a synthetic follow-up is stripped identically.
Baseline: 19 lesions, 34.98 ml. Classes follow Vanden Bulcke et al. 2025 bands: enlarging or shrinking beyond ±9 %/yr, stable within ±4 %/yr, trend between (1 trend up, 2 trend down, 1 stable). SEL candidates are the Elliott et al. 2019 Jacobian rule; a definite SEL is a cohort-relative ranking and needs a cohort.
The baseline scan of this subject was turned into a synthetic follow-up one modelled year later: the whole brain contracted by -2.0 %, 12 new lesions (29 to 1495 mm³) placed at least 6 mm from every existing lesion, a rigid repositioning (0.20, 0.30, -0.23 degrees; 0.52, -0.10, 0.52 voxels), gains 0.91, 1.01 and 2 % noise. Every baseline lesion moves with the field and nothing else; its truth volume in the synthetic mask is recorded per lesion, so each enlarging, shrinking or resolved call is judged against that lesion's own truth. The same pipeline then measured baseline against this synthetic scan.
| Report line | Truth | Pipeline reported | Reading |
|---|---|---|---|
| New lesions | 12 injected | 12 reported (12 new, 0 adjacent fragment); 0 false new, 0 false fragments; largest missed – | not graded (sensitivity 1.00) |
| False new lesions | 0 | 0 | not graded |
| Enlarging or shrinking calls on baseline lesions | 9 of 18 truly beyond ±9 %/yr | 0 false (0.0 %), 0 missed; class agreement 1.00 | read at face value |
| Resolved calls on baseline lesions | 0 of 18 | 0 | read at face value |
| SEL candidates on baseline lesions | 0 | 137 (7.21 per baseline lesion) | not trustworthy here |
| Brain volume change | -2.00 % | -2.52 % from the deformation field (recovery 1.26); -2.05 % from the mask ratio (recovery 1.03) | read with the calibration |
Injected lesions reported by size: <50 mm³: 3 of 3; 50-150 mm³: 3 of 3; 150-400 mm³: 2 of 2; >=400 mm³: 4 of 4.
Bands (mscard's own, stated so you can disagree): new-lesion sensitivity green ≥ 0.9, false new ≤ 1, brain-volume recovery [0.8, 1.2], false change rate ≤ 0.05, false resolved ≤ 0.0, SEL candidates per lesion ≤ 0.2; amber up to the second value of each.
Real pair, baseline T1 to T3: baseline lesions (cyan), new voxels on the warped follow-up (yellow), Jacobian expansion inside lesions with SEL candidate outlines (lime).
Calibration pair, baseline to synthetic follow-up, same rendering.
| Group | Class | Baseline mm³ | Follow-up mm³ | %/yr | Voxel centre (halfway grid) |
|---|---|---|---|---|---|
| 1 | trend_up | 30191 | 36061 | 8.9 | 49, 77, 102 |
| 6 | trend_down | 2962 | 2509 | -7.0 | 105, 141, 93 |
| 11 | enlarging | 363 | 590 | 28.6 | 123, 126, 94 |
| 14 | enlarging | 97 | 178 | 38.1 | 133, 107, 54 |
| 12 | shrinking | 236 | 116 | -23.2 | 119, 137, 90 |
| 7 | trend_down | 123 | 101 | -8.2 | 107, 143, 102 |
| 5 | shrinking | 223 | 91 | -27.0 | 76, 128, 131 |
| 13 | shrinking | 56 | 43 | -10.6 | 119, 95, 83 |
| 4 | shrinking | 68 | 35 | -22.2 | 72, 106, 123 |
| 17 | new | 0 | 18 | – | 72, 60, 103 |
| 19 | new | 0 | 18 | – | 76, 76, 109 |
| 16 | adjacent_fragment | 0 | 9 | – | 51, 108, 55 |
| 18 | new | 0 | 3 | – | 71, 160, 87 |
| 20 | new | 0 | 3 | – | 75, 81, 106 |
| 21 | adjacent_fragment | 0 | 2 | – | 90, 108, 94 |
| 2 | resolved | 29 | 0 | -45.7 | 59, 122, 96 |
| 10 | resolved | 221 | 0 | -45.7 | 117, 109, 87 |
| 9 | resolved | 32 | 0 | -45.7 | 114, 160, 86 |
| 8 | resolved | 74 | 0 | -45.7 | 112, 110, 81 |
| 15 | resolved | 78 | 0 | -45.7 | 136, 83, 102 |
1 stable lesions not listed; the full table is lesion_tracking.tsv.