When can multi-fragment 2D/3D registration be trusted? Calibrated certification and uncertainty-driven view planning for pelvic fracture fluoroscopy
Organizations: Department of Biomedical Engineering, Vanderbilt University, 2301 Vanderbilt Place, Nashville, TN 37235, USA
Abstract
Registering a segmented CT to fluoroscopy can locate every fracture fragment but can guide surgery only when trusted. After each shot, our closed loop registers all fragments, bounds their errors with a calibrated sandwich covariance, flags implausible or ambiguous ones, and stops or plans the next view. On simulated fluoroscopy of 60 held-out fracture CTs with realistic image uncertainty, 95% bounds covered 83% of unflagged fragments and 91% of those within 5mm, but from the second view on 14% failed by over 5mm, a third with bounds of at most 2mm. A fragment's shift when a view was added separated tightly bounded failures best (exploratory; AUC 0.94, bound 0.81). Requiring that the newest view moved no fragment beyond its bound, preferred post hoc, lowered false stops at 2mm from 23% to 6% overall, with 345 instead of 362 correct ones but more trajectories uncertified; delaying the stop did not, and other signals at equal acceptance left 10-15% false. With a segmentation error shared by all views (median Dice 0.89), 21% of this rule's stops were false against 10% without it (15 cases; exploratory). Minimax planning certified sooner than the clinical protocol without the check (4.20 versus 4.72 shots, Holm p <0.001), but no criterion reached the pre-specified endpoint, a true 2mm error, significantly sooner, even in the true basin. On real fluoroscopy of five intact pelves no 2mm certificate was false. Trust takes evidence across views, not only a calibrated local bound.
Figures & tables
| S1, failure: error 5 mm | S1, failure: error 2 mm | |||
| all | bound 2 mm | all | bound 2 mm | |
| Fragment registrations (failures) | 14864 (2153) | 10846 (679) | 14864 (2890) | 10846 (840) |
| AUC | ||||
| Prior distance | 0.84 [0.81, 0.86] | 0.91 [0.88, 0.94] | 0.77 [0.73, 0.80] | 0.87 [0.83, 0.90] |
| Ambiguity (objective gap) | 0.85 [0.83, 0.87] | 0.88 [0.86, 0.90] | 0.83 [0.81, 0.84] | 0.86 [0.83, 0.88] |
| Start disagreement | 0.83 [0.81, 0.85] | 0.87 [0.85, 0.89] | 0.82 [0.80, 0.84] | 0.85 [0.83, 0.88] |
| True | Local rule | Stable rule | |||||
| Setting | Policy | shots | shots | false stops | shots | false stops | correct |
| S1 | Minimax (ours) | 3.68 | 4.20 | 11/38 | 4.87 | 1/29 | 4.88 |
| A-optimal | 4.02 | 4.23 | 10/39 | 4.95 | 1/27 | 4.98 | |
| D-optimal | 3.50 | 4.32 | 6/41 | 4.77 | 1/37 | 4.80 | |
| E-optimal | 3.50 | 4.60 | 8/32 | 5.07 | 2/24 | 5.17 | |
| EIG (full) | 3.78 | 4.25 | 10/41 | 4.87 | 1/34 | 4.88 | |
| All policies (840 trajectories) | Minimax (60) | Protocol (60) | |||||||||
| false | false share | false share | |||||||||
| Stop rule | correct | 2–5 mm | 5 mm | none | [95% CI] | c / f / n | [exact CI] | c / f / n | false | ||
| Local (bound and flags) | 362 | 28 | 78 | 372 | 23% | [16, 29] | 27 / 11 / 22 | 29% | [15.4, 45.9] | 26 / 9 / 25 | 26% |
| no stop before shot 3 | 369 | 25 | 74 | 372 | 21% | [15, 28] | 29 / 9 / 22 | 24% | [11.4, 40.2] | 26 / 9 / 25 | 26% |
| two local certificates in a row | 294 | 7 | 24 | 515 | 10% | [6, 14] | 28 / 2 / 30 | 7% | [0.8, 22.1] | 18 / 2 / 40 | 10% |
| Stable (shift bound) | 345 | 8 | 15 | 472 | 6% | [4, 10] | 28 / 1 / 31 | 3% | [0.1, 17.8] | 21 / 3 / 36 | 12% |
| S1 | S5 | |
| Whole trajectories (15 cases, 105 per setting) | ||
| Coverage, unflagged | 80.0% | 76.7% |
| within 5 mm | 90.1% | 86.4% |
| Gross ( 5 mm), shot 2 | 19% | 18% |
| Local rule: c / f / n | 35 / 8 / 62 | 28 / 12 / 65 |
| f 2–5 / 5 mm | 1 / 7 | 4 / 8 |
| Predictor | coverage | held-out | coverage range | median | certified | |
|---|---|---|---|---|---|---|
| at | coverage | over volumes | bound (mm) | (false) | ||
| Sandwich, residual spectrum (ours) | 2.41 | 74% | 94% | 12 pts | 1.53 | 26 (6) |
| Sandwich, white noise | 0.826 | 96% | 94% | 12 pts | 1.55 | 23 (5) |
| Sandwich, white noise correlation area | 0.00118 | 100% | 94% | 8 pts | 1.68 | 23 (5) |
| Cramér–Rao, white noise | 1.49 | 88% | 93% | 15 pts | 1.58 | 21 (5) |
| Cramér–Rao, white noise correlation area | 0.00944 | 100% | 94% | 11 pts | 1.81 | 17 (4) |
| True 2 mm | Certified (local rule) | Correctly certified (stable rule) | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Setting | Baseline | diff. | 95% CI | Holm | diff. | 95% CI | Holm | diff. | 95% CI | Holm |
| S1 | A-optimal | 0.33 | [ 0.63, 0.05] | 0.179 | 0.03 | [ 0.22, 0.13] | 0.962 | 0.10 | [ 0.32, 0.10] | 1.000 |
| D-optimal | +0.18 | [ 0.13, 0.50] | 0.718 | 0.12 | [ 0.28, 0.07] | 0.405 | +0.08 | [ 0.12, 0.28] | 1.000 | |
| E-optimal | +0.18 | [ 0.15, 0.53] | 1.000 | 0.40 | [ 0.62, 0.20] | 0.004 | 0.28 | [ 0.47, 0.12] | 0.032 | |
| EIG (full) | 0.10 | [ 0.38, 0.18] | 1.000 | 0.05 | [ 0.17, 0.07] | 0.877 | +0.00 | [ 0.20, 0.18] | 1.000 | |
| EIG (diagonal) | +0.10 | [ 0.20, 0.38] | 1.000 | 0.28 | [ 0.47, 0.10] | 0.012 | 0.02 | [ 0.20, 0.15] | 1.000 | |
| Shots to true 2 mm | Oracle basin | ||||||
| Policy | deployable | oracle start | oracle basin | correctly | minimax policy (true) | ||
| certified | diff. | 95% CI | Holm | ||||
| Minimax (ours) | 3.93 | 4.23 | 3.07 | 4.67 | – | ||
| A-optimal | 4.27 | 4.20 | 3.03 | 4.67 | +0.03 | [ 0.47, 0.50] | 1.000 |
| D-optimal | 3.73 | 3.93 | 2.57 | 4.50 | +0.50 | [0.13, 0.90] | 0.160 |
| E-optimal | 3.60 | 3.83 | 2.67 | 4.93 | +0.40 | [ 0.03, 0.83] | 0.600 |
| Second view (after AP) | true 2 mm | true 1 mm |
| Minimax (ours) | 32% | 13% |
| A-optimal | 27% | 17% |
| D-optimal | 38% | 23% |
| E-optimal | 38% | 20% |
| EIG (full) | 30% | 23% |
| EIG (diagonal) | 33% | 13% |
| Specimen | mTRE (mm) | landmark (mm) | coverage | Shots to certification at 2 mm | ||
|---|---|---|---|---|---|---|
| median | median max | held-out | minimax | random | PolyPose | |
| 01 | 0.67 | 0.72 | 95% | 2.00 | 4.58 | 2.00 |
| 02 | 0.41 | 0.59 | 100% | 2.00 | 5.75 | 6.00 |
| 03 | 1.18 | 1.26 | 86% | – | – | – |
| 04 | 0.54 | 0.88 | 95% | 3.00 | 5.58 | 6.00 |
| 05 | 0.36 | 0.43 | 95% | 2.00 | 2.50 | 2.00 |
Appendix figures & tables17 assets
Supplementary material from the paper’s appendix.
Appendix
| Time (UTC) | Commit | Decision (decision log) |
| 2026-10-28 23:53 | 2f5c329 | Case split: development, calibration, test |
| 2026-10-29 13:55 | 921c82d | Realistic per-view uncertainty (redesign) |
| 2026-10-29 14:58 | 1d76d94 | Noise amplitude ( sensitivity) |
| 2026-10-29 16:47 | 1653bbd | Registration for realistic noise (tempering, search) |
| 2026-10-29 17:59 | 26f603d | Study protocol: settings, policies, endpoints, statistics |
| 2026-10-29 18:29 | b422ae6 | Calibration rule amended |
| True 2 mm | Local rule | Stable rule | |||||
| Setting | Policy | shots | shots | false/cert. | false traj. | shots | false traj. |
| S1 tracked, synthetic ( ) | Minimax (ours) | 3.88 | 4.47 | 16/148 | 20% | 5.20 | 3% |
| A-optimal | 4.32 | 4.55 | 20/138 | 20% | 5.40 | 2% | |
| D-optimal | 3.67 | 4.62 | 10/141 | 10% | 5.13 | 2% | |
| E-optimal | 3.72 | 4.98 | 10/118 | 13% | 5.52 | 3% | |
| EIG (full) | 3.95 | 4.53 | 14/147 | 18% | 5.20 | 2% | |
| Shots to true 2 mm | Shots to certification | ||||||
| Setting | Baseline | diff. | 95% CI | Holm | diff. | 95% CI | Holm |
| S1 tracked, synthetic | A-optimal | 0.43 | [ 0.82, 0.07] | 0.230 | 0.08 | [ 0.33, 0.15] | 0.657 |
| D-optimal | +0.22 | [ 0.13, 0.58] | 0.732 | 0.15 | [ 0.33, 0.05] | 0.140 | |
| E-optimal | +0.17 | [ 0.23, 0.57] | 1.000 | 0.52 | [ 0.80, 0.27] | 0.001 | |
| EIG (full) | 0.07 | [ 0.40, 0.27] | 1.000 | 0.07 | [ 0.22, 0.10] | 0.386 | |
| EIG (diagonal) | +0.17 | [ 0.17, 0.52] | 1.000 | 0.30 | [ 0.50, 0.08] | 0.005 | |
| Setting | Baseline | minimax | baseline | diff. | 95% CI | Holm |
|---|---|---|---|---|---|---|
| S1 | A-optimal | 5.20 | 5.40 | 0.20 | [ 0.47, 0.03] | 0.907 |
| D-optimal | 5.20 | 5.13 | +0.07 | [ 0.13, 0.28] | 1.000 | |
| E-optimal | 5.20 | 5.52 | 0.32 | [ 0.57, 0.08] | 0.115 | |
| EIG (full) | 5.20 | 5.20 | +0.00 | [ 0.23, 0.23] | 1.000 | |
| EIG (diagonal) | 5.20 | 5.22 | 0.02 | [ 0.23, 0.20] | 1.000 | |
| Gradient heuristic | 5.20 | 5.33 | 0.13 | [ 0.38, 0.10] | 0.907 |
| Setting | Value |
|---|---|
| Pose prior (per axis) | 5 ∘ , 10 mm |
| Working resolutions | 32 px, 64 px, 128 px |
| Adam step sizes (rotation, translation) | 0.01 rad, 1 mm (scaled by pixel size per resolution) |
| Iterations per stage (max.), plateau patience | 300, 10 |
| Trust region per stage (coarsest) | 30 ∘ , 20 mm (halved per finer resolution) |
| Starts | 2 (previous estimate, prior mean); selection by the polished MAP objective |
| True 2 mm | Certified (local rule) | Correctly certified (stable rule) | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Setting | Baseline | diff. | 95% CI | Holm | diff. | 95% CI | Holm | diff. | 95% CI | Holm |
| S2 | A-optimal | 0.33 | [ 0.80, 0.00] | 0.251 | +0.00 | [ 0.27, 0.27] | 1.000 | +0.00 | [ 0.40, 0.40] | 1.000 |
| Clinical protocol | +0.13 | [ 0.60, 0.93] | 1.000 | 0.60 | [ 1.00, 0.20] | 0.060 | 0.33 | [ 0.73, 0.00] | 0.251 | |
| Max separation | 1.00 | [ 1.67, 0.47] | 0.022 | 0.60 | [ 1.20, 0.07] | 0.160 | 0.53 | [ 0.93, 0.13] | 0.104 | |
| Random | 0.16 | [ 0.58, 0.27] | 1.000 | 0.42 | [ 0.80, 0.11] | 0.060 | 0.16 | [ 0.51, 0.13] | 1.000 | |
| S3 | A-optimal | +0.07 | [ 0.67, 0.73] | 1.000 | 0.07 | [ 0.27, 0.13] | 0.564 | 0.07 | [ 0.47, 0.33] | 0.922 |
| mm | mm | ||||||
| Setting | Policy | true | certified | false/cert. | true | certified | false/cert. |
| S1 | Minimax (ours) | 4.92 | 6.52 | 0/29 | 3.55 | 3.38 | 36/208 |
| A-optimal | 5.10 | 6.57 | 0/26 | 3.98 | 3.52 | 37/192 | |
| D-optimal | 4.65 | 6.60 | 0/24 | 3.48 | 3.48 | 26/207 | |
| E-optimal | 5.13 | 6.63 | 0/22 | 3.43 | 3.73 | 19/189 | |
| EIG (full) | 4.83 | 6.62 | 0/23 | 3.77 | 3.50 | 34/207 | |
| Shots to true | Shots to certification (local) | ||||||
| Setting | Baseline | diff. | 95% CI | Holm | diff. | 95% CI | Holm |
| S1 | A-optimal | 0.18 | [ 0.55, 0.20] | 0.978 | 0.05 | [ 0.13, 0.03] | 0.308 |
| D-optimal | +0.27 | [ 0.13, 0.65] | 0.978 | 0.08 | [ 0.20, 0.02] | 0.308 | |
| E-optimal | 0.22 | [ 0.65, 0.20] | 1.000 | 0.12 | [ 0.22, 0.03] | 0.072 | |
| EIG (full) | +0.08 | [ 0.28, 0.45] | 1.000 | 0.10 | [ 0.22, 0.00] | 0.280 | |
| EIG (diagonal) | 0.05 | [ 0.38, 0.27] | 1.000 | 0.18 | [ 0.33, 0.07] | 0.066 | |
| Shots to true | Shots to certification (local) | ||||||
| Setting | Baseline | diff. | 95% CI | Holm | diff. | 95% CI | Holm |
| S1 | A-optimal | 0.43 | [ 0.82, 0.07] | 0.391 | 0.13 | [ 0.43, 0.15] | 0.609 |
| D-optimal | +0.07 | [ 0.28, 0.43] | 1.000 | 0.10 | [ 0.27, 0.07] | 0.497 | |
| E-optimal | +0.12 | [ 0.30, 0.53] | 1.000 | 0.35 | [ 0.62, 0.08] | 0.032 | |
| EIG (full) | 0.22 | [ 0.58, 0.15] | 0.889 | 0.12 | [ 0.30, 0.07] | 0.336 | |
| EIG (diagonal) | +0.03 | [ 0.28, 0.37] | 1.000 | 0.22 | [ 0.47, 0.05] | 0.089 | |
| Setting | Volume (cm 3 ) | registrations | gross ( 5 mm) | coverage | coverage, non-gross |
|---|---|---|---|---|---|
| S1 | 0–15 | 2690 | 19% | 78% | 95% |
| 15–35 | 3981 | 8% | 89% | 96% | |
| 35–100 | 3007 | 15% | 76% | 89% | |
| 100 | 4199 | 1% | 86% | 88% | |
| S2 | 0–15 | 223 | 24% | 71% | 90% |
| 15–35 | 450 | 11% | 84% | 93% |
| Stratum | Endpoint | minimax | fixed | D-opt. | minimax fixed | minimax D-opt. | ||
|---|---|---|---|---|---|---|---|---|
| diff. | 95% CI | diff. | 95% CI | |||||
| 3 ( ) | true | 3.38 | 4.00 | 3.19 | 0.62 | [ 1.38, 0.15] | +0.19 | [ 0.31, 0.69] |
| certified | 3.54 | 4.35 | 3.77 | 0.81 | [ 1.19, 0.46] | 0.23 | [ 0.46, 0.00] | |
| =4 ( ) | true | 3.95 | 3.86 | 3.81 | +0.10 | [ 0.67, 0.90] | +0.14 | [ 0.48, 0.71] |
| certified | 5.24 | 6.00 | 5.24 | 0.76 | [ 1.14, 0.43] | +0.00 | [ 0.38, 0.43] | |
| 5 ( ) | true | 4.77 | 4.38 | 4.38 | +0.38 | [ 0.31, 1.08] | +0.38 | [ 0.54, 1.31] |
| S1 ( 5 mm) | S3 ( 5 mm) | S2 ( 5 mm) | ||||
| all | bound 2 mm | all | bound 2 mm | all | bound 2 mm | |
| Fragment registrations (failures) | 14864 (2153) | 10846 (679) | 1788 (169) | 1445 (63) | 1790 (495) | 1123 (153) |
| AUC | ||||||
| Prior distance | 0.84 | 0.91 | 0.88 | 0.90 | 0.88 | 0.90 |
| Ambiguity (objective gap) | 0.85 | 0.88 | 0.88 | 0.86 | 0.80 | 0.83 |
| Start disagreement | 0.83 | 0.87 | 0.86 | 0.85 | 0.82 | 0.86 |
| Rule | Case | Policy | Shot | cm 3 | Dice | Error | Bound | Shift | S1 error | Class |
|---|---|---|---|---|---|---|---|---|---|---|
| Stable | 019 | Random (2) | 4 | 43 | 0.87 | 6.46 | 1.70 | 1.07 | 6.26 | present in S1 |
| 031 | Minimax | 4 | 45 | 0.88 | 2.51 | 1.40 | 0.26 | – | path differs | |
| 031 | Random (2) | 4 | 45 | 0.88 | 4.75 | 1.39 | 1.35 | 5.91 | present in S1 | |
| 074 | Minimax | 5 | 40 | 0.87 | 28.22 | 1.53 | 1.43 | – | path differs | |
| 092 | Max separation | 4 | 32 | 0.77 | 2.06 | 1.93 | 0.53 | 1.16 | new, 2–5 mm | |
| 093 | Protocol | 5 | 22 | 0.85 | 3.63 | 1.77 | 1.64 | 3.40 | present in S1 |
| median error (mm) | error 2 mm | coverage | |||||
| Fragments | S1 | S5 | S1 | S5 | S1 | S5 | |
| Dice | |||||||
| 0.80 | 63 | 0.73 | 0.99 | 9.5% | 22.2% | 100.0% | 90.5% |
| 0.80–0.85 | 179 | 0.92 | 1.05 | 12.8% | 12.3% | 96.6% | 96.1% |
| 0.85–0.90 | 423 | 0.64 | 0.70 | 4.5% | 7.3% | 89.4% | 84.6% |
| 0.90–0.95 | 378 | 0.49 | 0.51 | 4.5% | 5.3% | 89.4% | 88.9% |
| S2 run | gross | coverage | coverage, | false/cert. | minimax | minimax fixed |
|---|---|---|---|---|---|---|
| 5 mm | unflagged | non-gross | (all policies) | certified | (Holm ) | |
| first run (deviation) | 30% | 65% | 82% | 23/105 | 5.53 | 0.87 (0.061) |
| re-run (as specified) | 28% | 76% | 88% | 8/86 | 5.47 | 1.00 (0.013) |
| Shots to true 2 mm | Oracle basin | ||||||
| Policy | deployable | oracle start | oracle basin | certified | minimax policy (true) | ||
| shots | diff. | 95% CI | Holm | ||||
| Minimax (ours) | 4.20 | 4.57 | 3.17 | 4.43 | – | ||
| A-optimal | 4.67 | 4.50 | 3.03 | 4.40 | +0.13 | [ 0.40, 0.63] | 1.000 |
| D-optimal | 3.93 | 4.27 | 2.60 | 4.47 | +0.57 | [0.13, 1.03] | 0.150 |
| E-optimal | 3.83 | 4.13 | 2.70 | 4.97 | +0.47 | [0.00, 0.93] | 0.572 |