△ Mechanism confirmed, baseline not beaten 2026

Boundary-Radial Persistence Loss

Implementation & benchmark of arXiv:2608.11963 — Computing extended persistent homology of radial distance filtrations of Euclidean shapes

Usefulness7/10

Difficulty6/10

Novelty6/10

Regularization Accuracy Generalization Memory Cnn Loss Regularization Geometry Topology Geometry Algebraic topology Computation

Source paper: Computing extended persistent homology of radial distance filtrations of Euclidean shapes arXiv:2608.11963 ⓘ · analyzed Sep 1, 2026

AI-generated research hypothesis, automatically tested. Not peer-reviewed.

Idea description

Add a topology-aware loss to a segmentation or implicit-shape network by computing radial extended persistence on the predicted boundary rather than on the full predicted mask. Match signed persistence intervals of the prediction to those of the target, penalizing missing, extra, or incorrectly ordered radial components and holes. This should provide a compact shape prior that is sensitive to anatomy-specific radial organization while avoiding volumetric homology computation.

Formulas

$$\rho_c(x)=\lVert x-c\rVert_2,\qquad A_\alpha=\left(\rho_c^A\right)^{-1}(-\infty,\alpha],\qquad A^\alpha=\left(\rho_c^A\right)^{-1}[\alpha,\infty).$$

$$\mathrm{XRPH}_k(A,c)=\bigoplus_{[b,d)\in S_k^A}\mathcal{I}_{[b,d)}.$$

$$\mathrm{XRPH}_k(M,c)=\mathrm{Ord}_k(M,c)\oplus\mathrm{Rel}_k(M,c)\oplus\mathrm{Ess}_k^+(M,c)\oplus\mathrm{Ess}_k^-(M,c),\qquad \mathrm{Ess}^+:s<t,\quad \mathrm{Ess}^-:t<s.$$$

$$\mathcal L_{\mathrm{rad}}=\min_{\pi}\sum_i w_i\left(\lvert b_i-\tilde b_{\pi(i)}\rvert+\lvert d_i-\tilde d_{\pi(i)}\rvert\right)+\lambda\,\#\mathrm{unmatched}.$$

Mathematical statement

For a center c and set A, the radial function is $\rho_c(x)=\|x-c\|_2$, with sublevel and superlevel sets $A_\alpha=\{x\in A:\rho_c(x)\leq\alpha\}$ and $A^\alpha=\{x\in A:\rho_c(x)\geq\alpha\}$. The degree-k extended radial persistence module decomposes as $\mathrm{XRPH}_k(A,c)=\bigoplus_{[b,d)\in S_k^A}\mathcal I_{[b,d)}$, where each interval records a birth radius b and death radius d. Ordinary classes have both endpoints in the ordinary part, relative classes have both endpoints in the relative part, and essential classes connect an ordinary birth $(s,\mathrm{Ord})$ to a relative death $(t,\mathrm{Rel})$, split into $\mathrm{Ess}^+$ when $s<t$ and $\mathrm{Ess}^-$ when $t<s$. For the planar boundary reduction stated in the paper, the interval decomposition of $\mathrm{XRPH}_0(M,c)$ and $\mathrm{XRPH}_1(M,c)$ is obtained from the signed annotation of $\mathrm{XRPH}_0(\partial M,c)$ and $\mathrm{XRPH}_1(\partial M,c)$ by retaining intervals whose birth has positive sign; when $c\notin M$, $\mathrm{XRPH}_2(M,c)=0$. Use a persistence matching cost such as $\mathcal L_{\mathrm{rad}}=\min_\pi\sum_i w_i(\lvert b_i-\tilde b_{\pi(i)}\rvert+\lvert d_i-\tilde d_{\pi(i)}\rvert)+\lambda\,\#\mathrm{unmatched}$, where predicted and target intervals are matched by $\pi$.

Implementation notes

Integrate this at the output of a binary or soft segmentation CNN, preferably after extracting a one-pixel predicted contour with a differentiable soft boundary score. Choose a fixed center c from the object centroid, an anatomical landmark, or a training-set convention. Convert contour pixels or a polygonal contour into a graph of vertices and edges; assign each vertex the radius r_i=||x_i-c||_2 and assign each edge the filtration value max(r_i,r_j), exactly matching the paper's simplex-wise radial filtration. Run a union-find persistence routine for H0 on the sublevel filtration and a reversed filtration for relative or superlevel events; in 2D, additionally compute H1 using a cubical-complex persistence library. Apply the paper's signed-birth rule to retain positive-birth intervals when reconstructing the interior radial classes from the boundary. Pseudocode: r=norm(vertices-c); f_edge=max(r[u],r[v]); bars=extended_persistence(complex,f_vertex,f_edge); matched=min_cost_match(pred_bars,target_bars); L=L_seg+alpha*L_rad. Initially stop gradients through the discrete matching and validate the scalar signal; then route gradients to birth/death contour pixels or use a soft persistence surrogate. The paper supplies the radial objects and boundary reduction; the engineer estimates c, extracts the contour, and computes the matching empirically. Test a small U-Net on synthetic circles, annuli, and branching shapes against Dice plus boundary loss. Success means fewer topological errors at equal Dice and lower peak memory than volumetric persistence.

Verification

Mechanism confirmed, baseline not beaten

Stage 1 · Toy mechanism gate: Passed ✓

Stage 2 · Mechanism transferred to benchmark: Confirmed ✓

Stage 2 · Practical benchmark result: No measurable effect

Methodology: 8 paired seeds, permutation test (p<0.05) against a learning-rate-tuned baseline. How verification works

Stage 1 — Mechanism check agent confidence 8/10

Built a readable boundary-radial persistence MVP with polygonal boundary H0 filtration, capped intervals, Hungarian endpoint matching, and unmatched penalties. The mechanism checks matched the quantitative predictions exactly: identical barcodes gave zero loss, a uniform perturbation of two radial intervals produced the predicted slope 4.0, and q unmatched components produced exactly qλ. A raster proxy assigned a penalty of 10.0 to an extra component despite Dice 0.938, versus 0.260 for a shifted single component with Dice 0.865; this is a promising topology-sensitive signal, but not evidence of CNN accuracy improvement.

Agent confidence: 8/10

Baseline: Raster proxy endpoint L1 on fixed noisy radii: 0.172234; shifted-mask Dice 0.8646; extra-component Dice 0.9380.
Idea: Mechanism verification: zero-loss 0.0; observed perturbation slope 4.000000 vs predicted 4.0; maximum perturbation fit error 8.9e-16; unmatched-loss maximum error 0.0. Raster radial matching loss (lambda=10): shifted 0.2598, extra component 10.0.

Limitations: The MVP does not implement full extended radial persistence, relative/essential signed classes, H1, the paper's cascading center-inside cases, differentiable gradients, or a trained segmentation CNN. The quantitative checks use canonical radial intervals for concentric circular components, and the raster comparison summarizes each boundary component by median radius; therefore no claim of end-to-end segmentation improvement or memory reduction was tested.

How to run: python3 run_experiment.py && /home/maxwelhelp/main/bin/python3 mini_segmentation.py

Stage 2 — Benchmark vs. tuned baseline (arXiv:2608.11963) Sep 2, 2026

No significant difference vs. the tuned baseline (0 metric points, p=1, wins 0 / 0 paired seeds).

Benchmark: Radial_boundary_shapes
Model: local_mlp_shared
Paired seeds: 8
Baseline mean: 0
Idea mean: 0
Effect (Δ): 0 (n/a; negative = idea better)
Wins: 0 / 0 paired seeds
p-value: 1 (permutation test, 20 000 shuffles)
Smallest detectable effect: n/a
Mechanism: Confirmed ✓
Practical effect: No effect
Baseline tuning: swept over 0 configs

Limitations:

This is a compact radial-boundary regression proxy rather than full raster segmentation. Full extended persistence, relative/essential classes, H1, differentiable contour extraction, volumetric memory, and CNN segmentation were not tested. The training intervention uses a differentiable sorted-radius surrogate rather than exact discrete Hungarian matching.

How to run: python3 bench_radial.py

Verdict computed by deterministic test code from paired-seed statistics — not by the language model.

Stage 2 — Benchmark Sep 1, 2026 latest Failed ✗

Agent confidence: 9/10

Built a structurally matched custom radial-boundary regression track with identical MLP systems for baseline MSE and the radial matching intervention. Across 8 paired seeds, the best baseline mean test MSE was 0.018877, while positive radial-loss weights worsened performance; the best idea setting was alpha=0 and therefore identical to baseline. The trained-model mechanism signature was confirmed by a positive MAE/radial-loss correlation of 0.878, but there was no significant task-metric win.

Baseline: lr=0.01, alpha=0.0; mean test MSE 0.0188770521
Idea: Best lr=0.01, alpha=0.0; mean test MSE 0.0188770521. alpha=0.05 mean 0.0210358966; alpha=0.2 mean 0.0285936717. Paired delta 0.0, permutation p=1.0.

Benchmark result