PaperlayerView exact revision on arXiv

Source-linked AI summary

Are Learned Vascular Tree Generators Better Than Conditional Resampling?

Reza Akbarian Bafghi, Sukirt Thakur, Maziar Raissi

arXiv:2610.30449v1cs.CE

TL;DR

Fidelity scores compare generated vascular trees with held-out real trees, so they cannot by themselves reveal reuse of training data. The paper evaluates a training-free conditional resampler against learned generators and proposes a branch-record near-copy diagnostic. The resampler leads the learned baselines on the reported fidelity comparisons, yet about half its branch records are near-copies, motivating reporting the diagnostic with controls.

  • Problem

    Fidelity scores compare generated trees with held-out trees rather than training data, so they cannot establish whether generated trees reuse training records.

    Source: §1 · p. 2

  • Method

    The paper compares learned generators with a conditional resampler and measures copying through branch records, reporting near-copy rates alongside fidelity scores and controls.

    Source: §1 · p. 3, §3.1 · p. 4, §3.4 · p. 6

  • Results

    The resampler leads three retrained generators on the reported fidelity scores, with exceptions against VesselGPT on IntrA and scale-free features, while about half its branch records are near-copies.

    Source: §1 · p. 2, §7 · p. 12

  • Takeaways & Limitations

    Near-copy rate is a diagnostic to report beside fidelity with a training copy, a jittered copy, and real held-out trees as controls.

    Source: §1 · p. 3, §7 · p. 12

    The learned-baseline comparison is bounded by the training budgets and protocol, including early stopping, an IntrA anchor-check failure for VesselVAE, and fidelity loss from PartVessel preprocessing.

    Source: §3.3 · p. 6

Abstract

from arXiv · show

Fidelity scores compare generated vascular trees with held-out real trees, so by themselves they cannot detect reuse of training data. A training-free conditional resampler shows this. On coronary trees (ImageCAS) and on IntrA, about half of its branch records (branches with their parent context) are near-copies of training records. Yet on our two fidelity scores, 1-NNA on tree summaries and a parent-daughter dependence AUC, it has better point estimates than three learned baselines as retrained here. A training copy scores better still. Adding noise to the copied values traces a curve of fidelity against the measured branch-record near-copy rate. At their own coronary rates (at most 1.3%), two learned baselines lie above this curve (worse) on parent-daughter relations. VesselGPT is inconclusive against the curve, as it is against the resampler on IntrA and on scale-free tree features. The curve is a reference, not a bound: real validation trees lie below it on both coronary scores, so low near-copy rates are not a goal in themselves. Jittered training copies also lie below it. On coronary data, one with a near-copy rate of about 1% does so on the dependence AUC, and so does one whose stronger jitter ruins its tree summaries. On IntrA, a lightly jittered copy lies below it on 1-NNA. A matched near-copy rate therefore certifies neither score against perturbed copies. The study is exploratory. We offer the near-copy rate as a diagnostic, reported beside fidelity with three controls: a training copy, a jittered copy and real held-out trees.

1 Introduction

Fidelity scores can reward training-data reuse because they compare generated trees with held-out trees, not the training set. The paper compares a conditional resampler with retrained learned generators and proposes measuring reuse at the branch-record level.

1 Introduction

  • The training-free resampler grows trees by copying daughters from training bifurcations with similar parent branches, providing a reference for learned generators.The illustrated tree combines daughter pairs from three different patients, and each copied branch length is within 0.45 mm of its source.Source: §1 · pp. 1–2, Figure 1 · p. 2
  • Fidelity alone cannot establish novelty: a training copy and real validation trees have the same coronary 1-NNA despite sharply different near-copy rates.The training copy and real validation trees both score 0.545, at 100% and 6.8% near-copies, respectively.Source: §1 · p. 2, Figure 2 · p. 3
  • The proposed diagnostic counts near-copied branch records, measuring reuse at the local bifurcation level rather than through whole-tree similarity.A branch record includes a branch with its parent and closest sibling; generated records are counted when closer to a training record than nearly all real held-out records.Source: §1 · pp. 2–3
  • The resampler leads the learned baselines on coronary fidelity scores and most IntrA comparisons, while about half its branch records are near-copies of training records.Its lead is not detected against VesselGPT on IntrA or on scale-free tree features; the comparison against learned baselines is exploratory and covers two fidelity scores.Source: §1 · pp. 2–3
  • At matched coronary near-copy rates, VesselVAE and PartVessel lie above the resampler’s noise-traced curve on parent-daughter relations, while VesselGPT is inconclusive.The curve is a reference rather than a bound: held-out real trees and perturbed training copies also lie below it.Source: §1 · pp. 2–3
  • The paper proposes reporting near-copy rates beside fidelity scores, with training-copy, jittered-copy, and real held-out-tree controls.The contribution also contrasts this branch-level diagnostic with a whole-tree data-copying test that does not detect copied branch records.Source: §1 · p. 3

2 Related work

Prior work studies copying, example-based synthesis, and vascular-tree generation, but existing fidelity measures are not designed to distinguish fidelity from copying. This paper measures reuse per branch record and compares learned generators with a resampling reference.

2 Related work

  • The paper’s novelty is measuring copying per branch record and comparing learned baselines with a resampler at the same measured near-copy rate.Source: §2 · p. 3
  • Prior copying work includes memorisation tests for point clouds, authenticity scoring for resampled outputs, and demonstrations that image diffusion models can reproduce training images.The cited examples include Achlioptas et al. [2018], Alaa et al. (2022), Carlini et al. [2023], Somepalli et al. [2023], and Feldman et al. [2025].Source: §2 · p. 3
  • The data-copying test compares generated and held-out distances to the nearest training point; the near-copy rate is proposed as a diagnostic alongside fidelity and controls, not as a better test.Source: §2 · p. 4
  • Example-based synthesis and smoothed resampling provide precedents for copying context-matched pieces and adding noise to resampled values.The paper adapts these ideas by assigning each branch the daughters of a similar training parent, using the resampler as a reference rather than a proposed generator.Source: §2 · p. 4
  • Existing vascular-tree generators use varied shape and point-cloud scores, none designed to separate fidelity from copying, and rankings under this paper’s scores need not transfer to point-cloud scores.Source: §2 · p. 4

3 Setup

The setup compares a training-free conditional resampler with three learned baselines on coronary and IntrA trees, using shared patient splits and fidelity and copying measures. The resampler copies daughters from similar training parents, with noise controlling how closely it copies.

3 Setup

  • The comparison uses a training-free resampler, three retrained learned baselines, shared patient splits, and a copying score beside fidelity.The resampler serves as a reference that copies by design.Source: §3 · p. 4, §3.1 · p. 4

3.1 The resampler

  • Each open branch copies the daughters of a similar training parent; noise perturbs copied continuous values and sweeping its level traces fidelity against copying.The resampler uses the training set and K rather than optimised parameters.Source: §3.1 · pp. 4–5, Figure 3 · p. 5

3.2 Data and compared methods

  • Coronary trees come from ImageCAS, while IntrA contributes 785 trees from 72 patients in five patient-grouped folds; controls include training copies, jittered copies, and real validation trees.The three learned baselines are VesselGPT, VesselVAE, and PartVessel, retrained from the authors’ code.Source: §3.2 · pp. 5–6

3.3 Training budgets of the learned baselines

  • The learned-baseline comparison is bounded by short training runs, a failed VesselVAE IntrA anchor check, and fidelity loss from PartVessel preprocessing.The authors state that verdicts describe these retrains, not the methods.Source: §3.3 · p. 6

3.4 Metrics

  • Fidelity is measured with whole-tree 1-NNA and a dependence AUC that emphasizes parent-daughter relations; closer to 0.5 is better on both scores.The dependence AUC classifier distinguishes generated from real branch records after per-class marginal normalization.Source: §3.4 · p. 6
  • The near-copy rate counts generated branch records closer to training records than all but q% of real test records; with q = 5, about half of resampler records qualify, versus 1.3% of VesselGPT records.The rate is reported beside fidelity to flag excess proximity to training records.Source: §3.4 · pp. 6–7, Figure 4 · p. 7

4 The leaderboard

The training-free resampler has the best point estimates among the generators on both fidelity scores across both datasets, though the training-copy control scores better still. Its results reflect copied branch descriptors and parent context, and the measured advantage depends partly on the evaluation features.

4 The leaderboard

  • Even on 11 whole-tree summaries, the resampler matches real trees best among generators despite conditioning only on one parent’s key.Its copied daughter pairs carry real branch lengths, angles, and parent-daughter relations; matching the parent is what makes those relations realistic.Source: §4 · p. 7
  • The resampler has the best point estimate on both fidelity scores on both datasets, while the training copy scores better still.On coronary data its lead over every learned baseline is detected on both scores; on IntrA, detected leads include PartVessel on both and VesselVAE on 1-NNA, while VesselGPT comparisons are inconclusive.Source: §4 · pp. 7–8, Figure 5 · p. 8
  • The resampler’s point-estimate lead is partly built in because it copies branch quantities used to compute both fidelity scores.For scale-free tree features, no 1-NNA difference from VesselGPT is detected, and VesselGPT’s point estimate is closer to 0.5.Source: §4 · p. 8

5 What the leaderboard cannot see

The resampler creates new whole trees by recombining branch records, but about half of those records are near-copies of training records. The fidelity scores do not establish novelty, motivating a branch-record near-copy diagnostic alongside them.

5 What the leaderboard cannot see

  • About half of the resampler’s branch records are near-copies of training records: 51.7% on coronary trees and 57.3% on IntrA.Learned baselines stay at or below the 5% rate of real held-out branch records; real validation trees reach 6.8%, so a lower rate is not inherently better.Source: §5 · p. 8
  • Near-copying is concentrated at the bifurcation scale: a whole-tree version of the near-copy rule flags only 6.3–7.2% of resampler trees.The rate rises as the copied daughter pair’s parent more closely matches its new parent.Source: §5 · p. 9
  • Fidelity scores cannot establish novelty: the training copy has the best 1-NNA and dependence AUC on both datasets.The scores compare generated trees with held-out trees, so a training copy can still resemble a new patient from the same population.Source: §5 · p. 9
  • Across the coronary sweep, reducing the near-copy rate from 82% to 14% barely moves either fidelity score, with confidence intervals overlapping.The scores appear to reflect realistic branch values and parent-daughter relations rather than copying as such.Source: §5 · p. 9
  • Depth-only conditioning lowers the coronary near-copy rate to 9.5% but worsens dependence AUC to 0.768, versus 0.613 for the curve’s λ = 0 point.The parent context that makes records near-copies also makes parent-daughter relations realistic.Source: §5 · p. 9
  • The whole-tree copying test does not flag the resampler, while its branch-record version reuses the near-copy rate’s records and may be anti-conservative.The study therefore offers the near-copy rate as a diagnostic, not a complete measure of reuse.Source: §5 · p. 10

6 Comparing at the same near-copy rate

At matched branch-record near-copy rates, no learned baseline is detectably closer to real trees than the conditional resampler. But real trees and perturbed copies also outperform the resampling curve, so it is a reference rather than a fidelity bound.

6 Comparing at the same near-copy rate

  • Sweeping resampler noise traces fidelity against near-copy rate; each learned model is compared with the curve’s score at its measured rate, with the gap matching only thresholded branch-record proximity.A negative gap means the model is closer to real trees than rate-matched noisy resampling; other forms or amounts of reuse are not ruled out.Source: §6 · p. 10

Reading Figure 7

  • The resampler curve is initially flat: near-copies can be removed almost for free at K = 30 or λ = 0.25 (14% retained), but fidelity worsens quickly at learned baselines’ low rates.Part of the steep rise may reflect leaving the data’s support; at λ = 3, 0.68 trees are valid on coronary validation runs.Source: §6 · p. 10

The learned baselines

  • No learned baseline lies below the resampler curve at its own near-copy rate; two are above it on coronary dependence AUC, while VesselGPT is inconclusive.PartVessel is also above the curve on coronary 1-NNA; on IntrA, VesselGPT is below it on 1-NNA in 2 of 5 folds.Source: §6 · p. 10, Figure 7 · p. 11
  • Coronary gap-to-curve intervals have half-widths of about 0.04 to 0.07, so gaps of that size can go undetected.The independent sampler lies above the curve on coronary dependence AUC, showing the comparison catches a large defect but is not a power analysis.Source: §6 · pp. 10–11

Perturbed copies lie below the curve

  • A training-tree copy with 0.5 mm Gaussian point jitter has a 0.83% coronary near-copy rate yet lies below the curve on dependence AUC and is never above it on IntrA.A 1 mm jittered copy also lies below the curve on dependence AUC but has 1-NNA 0.998; on IntrA, quarter-point-spacing jitter lies below the curve on 1-NNA in all 5 folds.Source: §6 · p. 11

Real trees lie below the curve too

  • Real validation trees lie below the curve on both coronary scores, so the curve does not bound attainable fidelity and low near-copy rates are not goals in themselves.Their coronary 1-NNA is 0.545, versus about 0.70 for the resampler at the same near-copy rate; size-matched real trees are below the curve on IntrA 1-NNA in all 5 folds.Source: §6 · p. 11

7 Discussion

The resampler beats three retrained generators on the reported fidelity scores, yet reuses about half its branch records; training copies score better still. The near-copy rate exposes reuse, but matching it does not validate fidelity against perturbed copies, so it should be reported with controls rather than treated as a target.

7 Discussion

  • The resampler leads three retrained generators on the reported fidelity scores, but about half its branch records are near-copies and a training copy scores better still.The near-copy rate, unlike the whole-tree copying test, exposes this reuse.Source: §7 · p. 12
  • At matched near-copy rates, VesselVAE and PartVessel are worse on parent-daughter relations, while VesselGPT is inconclusive.The study also finds VesselGPT inconclusive against the resampler on IntrA and scale-free tree features.Source: §7 · p. 12
  • Perturbed training copies can lie below the fidelity-versus-copying curve, so a matched near-copy rate certifies neither score against them.Real held-out trees also lie below the curve.Source: §7 · p. 12
  • Report the near-copy rate beside fidelity scores with a training copy, a jittered copy, and real held-out trees as controls.The study is exploratory, and copying may be useful when users know trees recombine training anatomy; only meshing was tested.Source: §7 · p. 12

A Scope of the evidence

The evidence is exploratory and descriptive: test sets informed the resampler and evaluation, and the comparisons concern these data and retrains rather than the published methods. Results vary across real-tree and perturbed-copy controls, while training, preprocessing, and comparison choices qualify the findings.

A Scope of the evidence

  • The test sets informed resampler and evaluation design, leaving no untouched hold-out or confirmatory test; verdicts are descriptive, with no multiplicity correction across analyses.The comparison covers two datasets of small trees and these retrains, not the methods.Source: A Scope of the evidence · p. 15
  • The resampler has built-in advantages because it copies branch quantities used by both scores, while IntrA outputs binary trees and PartVessel is scored against the full fold training set.The resampler is fit per artery and scored on more coronary trees, whereas learned checkpoints and decoding options were validation-selected.Source: A Scope of the evidence · p. 15
  • Perturbed copies show that matching near-copy rates certifies neither dependence AUC on either dataset nor IntrA 1-NNA against perturbed copies.No tested jittered or smooth copy lies below the curve on scale-free 1-NNA.Source: A Scope of the evidence · pp. 15–16
  • The coronary near-copy thresholds and rates were stable under point re-sampling, but verdicts were not recomputed at the re-sampled thresholds.On coronary data, the validity flag excludes branch collisions, whose share rises from 0.11 to 0.42 as λ increases; on IntrA it does not measure generator validity.Source: A Scope of the evidence · p. 16
  • At q = 5, real coronary validation trees lie below the curve by about 0.16 on 1-NNA and 0.13 on dependence AUC; IntrA 1-NNA is below in 5 of 5 folds.IntrA dependence AUC on all features is below in 2 of 5 folds and otherwise not distinguished; no feature-subset fold is below.Source: A Scope of the evidence · p. 16

B.1 Budgets and stopping

  • VesselGPT and VesselVAE stopped far short of their authors’ default epoch counts, and the VesselVAE anchor check failed.No claim is made that the evaluated configurations faithfully reproduce the published models.Source: A Scope of the evidence · pp. 16–17
  • Validation losses stopped improving early while training losses kept falling, so selected checkpoints were not limited by the training budget on these criteria.VesselGPT and PartVessel stage 1 stopped by early stopping in every run.Source: §B.1 · p. 17

B.2 Departures from the authors’ code

  • The reported runs used rescaling and code changes, including re-seeded VesselVAE sampling and clamped PartVessel log-variance.VesselVAE checkpoints were selected before divergence.Source: §B.2 · p. 17

B.3 Other configurations, anchors and round-trip references

Other configurations scored worse than the reported ones, and the VesselVAE anchor check failed; PartVessel preprocessing alone also costs fidelity. The comparison has advantages on both sides, and loss curves show early validation stopping while training losses continue to fall.

B.3 Other configurations, anchors and round-trip references

  • Other configurations of all three learned baselines scored worse than the reported configurations on the coronary test set.Their reported versus other-configuration 1-NNA scores were 0.776 versus 0.984 for VesselGPT, 0.966 versus 0.997 for VesselVAE, and 0.999 versus 1.000 for PartVessel.Source: §B.3 · p. 17
  • PartVessel preprocessing alone costs fidelity, while the VesselVAE anchor check fails and its score describes the checkpoint as run in this evaluation.Real test trees passed through PartVessel preprocessing score 1-NNA 0.855 on coronary data and 0.666 ± 0.008 on IntrA.Source: §B.3 · p. 18

B.4 Fairness in both directions

  • The resampler’s per-artery fitting and larger coronary sample favor it, while validation-selected checkpoints and decoding options favor learned baselines.Validation losses stop improving early while training losses keep falling, so selected checkpoints are not limited by the training budget on those criteria.Source: §B.4 · p. 18, Figure 8 · p. 19

C The dependence test and the decision rules

The dependence test is calibrated against controls and reports AUC only when calibration and shuffle-sanity gates pass. Its decision rules require paired uncertainty, effect-size, and multiplicity checks, with all five IntrA folds needed to count a baseline as beaten.

C The dependence test and the decision rules

  • On validation data, two real samples score 0.53 [0.45, 0.61], while a column-permuted positive control scores 0.91 and shuffle sanity is 0.49–0.50.The test responds to one-hop conditioning: the resampler scores 0.58–0.59 for K from 1 to 30 and 0.76 with depth-only conditioning.Source: C The dependence test and the decision… · p. 20
  • A dependence AUC is reported only when the calibration gate passes and shuffled-record sanity AUC is at most 0.55.The gate passes on coronary test data and in 5/5 IntrA folds; VesselVAE fails the shuffle-sanity criterion on IntrA.Source: C The dependence test and the decision… · p. 20
  • A baseline counts as beaten on IntrA only when all 5 folds satisfy the paired decision rules, including Holm-adjusted significance and the noise threshold.The 1-NNA rule uses a paired bootstrap interval and Holm-adjusted p-value; IntrA equivalence additionally requires both pooled-bootstrap and across-fold intervals within ±0.045.Source: C The dependence test and the decision… · p. 20

D Reproducibility details

The appendix documents evaluation and resampling procedures, then gives full comparison tables and checks on copying, fidelity, and uncertainty. The resampler’s branch-record copying is detected even when its whole-tree result is inconclusive.

D Reproducibility details

  • The curve comparison uses paired uncertainty intervals and flags ties or extrapolations; its intervals condition on stored generated sets and exclude interpolation error.The method also floors near-copy rates for plotting and interpolation, so zero observed copies do not establish a zero population rate.Source: D Reproducibility details · p. 21
  • The resampler has the best fidelity point estimates among generators on both datasets, while about half its branch records are near-copies; the training copy scores better still.VesselVAE’s IntrA dependence AUC is unreported because it fails the shuffled-record sanity check.Source: Table 1 · p. 22
  • The whole-tree copying test does not flag the resampler, while its branch-record version does; CT magnitudes cannot be compared across levels or datasets.The whole-tree coronary result is underpowered, and the IntrA calibration check is undefined in one fold.Source: Table 3 · p. 23
  • At q = 5, no learned baseline lies below the curve on either dataset; VesselVAE and PartVessel lie above it on the coronary dependence AUC.IntrA verdicts are aggregated across five folds, and VesselVAE’s dependence AUC is not reportable.Source: Table 4 · p. 23
  • On IntrA, VesselGPT’s mean 1-NNA lies below the curve, but no learned baseline lies below it across all five folds.The plotted fold means have no intervals; only the training copy’s 1-NNA (0.494) lies below 0.5.Source: Figure 9 · p. 23
  • Near-copy rates are about 1%, 5%, and 10% for real test branch records by construction, while every learned baseline’s rate is below the real-record rate.The resampler’s near-copy rate is far above the real test-record rate at every threshold.Source: Table 7 · p. 24

E Full tables

The full tables compare fidelity, dependence, copying, and perturbed-copy controls. They show that resampler advantages and curve comparisons vary by dataset, measure, and copying perturbation.

E Full tables

  • On coronary data, no learned baseline lies below the curve at any threshold, while VesselVAE and PartVessel lie above it on dependence AUC.At q = 1, several comparisons are flagged at the rate-resolution floor, and no 1-NNA comparison passes the equivalence test.Source: Table 13 · p. 28
  • On IntrA, no learned baseline lies below the curve in all five folds, and no equivalence test passes.Each fold has its own curve; VesselVAE’s dependence AUC is not reportable.Source: Table 14 · p. 29
  • Real validation trees lie below the curve on both coronary scores, and a coronary jittered copy lies below it on dependence AUC.The IntrA real-validation 1-NNA comparison is size-matched and uses a curve recomputed at that size.Source: Table 15 · p. 30
  • On coronary data, a training copy with 1 mm jitter lies below the curve on dependence AUC, while pooled and per-artery resampler scores show no detectable difference.On scale-free features, no 1-NNA difference from VesselGPT is detected.Source: Table 16 · p. 31
  • On IntrA, a training copy jittered by a quarter of the point spacing lies below the curve on 1-NNA in all five folds.At q = 10, a 0.2-unit jittered copy lies below the curve on dependence AUC in all five folds, at the near-copy floor.Source: Table 17 · p. 32

F Meshing (coronary, n = 20 per method)

The meshing check shows that resampled coronary trees can be meshed, without establishing a validity ranking. IntrA gallery examples contrast learned-generator outputs with real trees.

F Meshing (coronary, n = 20 per method)

  • 14/20 resampler trees and 2/20 VesselGPT trees met the fixed meshing criterion, but the under-sized study claims no validity ranking.Real test trees: 11/20; the resampler-versus-VesselGPT Fisher exact test gives p = 2.4 × 10^-4.Source: F Meshing (coronary, n = 20 per method) · p. 33

G Galleries

  • Random IntrA draws show tangled VesselVAE trees and thin PartVessel trees unlike real trees.The displayed trees use a fixed seed, no selection, and the same renderer as Figure 6.Source: Figure 10 · p. 34

H Specification of the evaluation and the resampler

The appendix specifies how trees and branch records are normalized, how the conditional resampler generates trees, and how scores and uncertainty are computed. It also states the scope of validity checks and inference intervals.

H.1 Tree features and branch records

  • Tree processing parses branch polylines and parent-child edges, normalizes topology, and resamples branches at fixed arc-length spacing before computing tree features and branch records.A branch record pairs each non-root branch with its parent and closest sibling.Source: §H.1 · p. 35, Table 20 · p. 36

H.2 The resampler

  • The resampler draws a root record, generates branch paths matching target length and tortuosity, then selects nearby training bifurcations to supply daughter payloads.It uses a calibration table to choose path curvature and applies branch-count and depth caps.Source: §H.2 · pp. 36–37
  • A tree is valid only if it is connected and acyclic with one root, finite values, and at least two points per branch; the flag misses branch collisions.On coronary validation runs, valid-tree share falls from 0.98 at λ = 0 to 0.68 at λ = 3.0; IntrA scores are uninformative because real trees also have single-point branches.Source: §H.2 · p. 38

H.3 Inference procedures

  • Inference uses patient-level resampling, grouped cross-validation, and paired comparisons; reported intervals condition on stored generated sets and exclude interpolation error.The dependence AUC uses repeated classifier fits, while the curve comparison interpolates between stored noise levels.Source: §H.3 · pp. 38–40

H.4 Data

The study uses in-house coronary centerline trees from ImageCAS and patient-level splits; IntrA folds likewise keep each patient’s trees out of both training and test.

H.4 Data

  • Coronary data comprise 1,092 left and right centerline trees with radii from 546 ImageCAS patients, split by patient into 446 / 50 / 50 patients for training, validation and test.Each branch is a polyline of (x, y, z, r) points in millimeters; the trees were pruned, and the derived trees are not redistributed.Source: §H.4 · p. 40
  • IntrA’s 785 trees from 72 patients are assigned to folds by patient, with seven validation patients drawn from other folds and the remaining patients used for training.Each fold is a test set; training patients are those not listed for that fold, so no patient’s trees occur in both training and test.Source: §H.4 · p. 40, Table 22 · p. 41
  • The evaluation and resampler can be reimplemented from the stated definitions, constants and splits, but the coronary trees are not distributed.Table 22 lists the IntrA patients and training / validation / test tree counts for each fold.Source: §H.4 · p. 40, Table 22 · p. 41

Related papers

Loading 2610.30449v1…