{"id":"39e6e143-3375-42e6-af4e-38edfd1ed5e6","arxiv_id":"2607.22053","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A softmax + Cahn–Hilliard variational model with a mixed L2–H^{-1} flow separates homogeneous adjacent structures at weak boundaries and is proven well-posed with an unconditionally stable SAV-FFT solver.","lead":"The paper combines softmax region fitting with Cahn–Hilliard phase-separation dynamics into a variational model for separating adjacent structures that look alike and are divided only by weak boundaries. It contributes well-posedness proofs, a fast FFT-based unconditionally stable solver, and a 10-image SKI10 comparison showing training-free Dice 0.982 with the most stable HD95 among compared methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The distance-prior D_a in Eq. (8) is never specified; it can break label symmetry exactly where (I−c_a)² degenerates, so the claimed CH-driven separation may be an artifact of an unstated spatial prior.","rationale":"The reader's weakest assumption identifies exactly the gap I consider most load-bearing: the distance-prior D_a is introduced as part of the data term but its construction is never specified, and it is capable of doing the work attributed to Cahn–Hilliard regularization. The mathematical analysis in §3 and the SAV scheme in §4 appear internally consistent, and the ablation in Figure 6 does give some support to the claim that CH regularization helps preserve weak interfaces when the data force is weak. However, that ablation does not isolate CH from D_a, because both the ablated and complete models include the same distance prior. The paper's claim of 'training-free, competitive with supervised deep learning' is therefore contingent on D_a not encoding privileged spatial information. This is an empirical/fairness concern rather than a mathematical refutation, so it warrants a conditional verdict rather than rejection. The test I propose — running the SKI10 experiments with D_a disabled — would settle whether the observed separation of homogeneous structures is due to phase separation or to the unstated prior. I also note the abstract's 'improved boundary localization' is contradicted on mean HD95 by the paper's own Table 3, though the smaller standard deviation is a legitimate stability claim. Because the reader already reached CONDITIONAL and my concern is the same one, the verdict should remain unchanged pending the requested specification and ablation.","tokens_in":21870,"tokens_out":4890,"duration_ms":48186,"concrete_test":"Ask the authors to specify the exact construction of D_a and the values of γ_a, and to release code and D_a maps. Then run the decisive ablation: set γ_a = 0 (or D_a ≡ constant) for all phases on the same 10 SKI10 test images and report Dice, IoU, and HD95. If performance collapses or adjacent structures merge, the reported separation is driven by the unstated spatial prior, not by Cahn–Hilliard phase separation. As a complementary check, give CV or RSF the same D_a as a spatially varying weight and compare; if a classical model with the same prior matches or beats Ours, the CH contribution is not the active ingredient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central experimental claim is that the Cahn–Hilliard term supplies the missing phase-separation mechanism when the image-driven force degenerates near weak boundaries (§2.2, Figures 1–2). But the full model includes a per-phase distance-prior D_a in the data fidelity term: E_data = λ Σ_a ∫ Φ_a(U)[(I−c_a)² + γ_a D_a] dx (Eq. 8). The paper states only that D_a 'encodes spatial prior information by assigning smaller values to pixels closer to the expected support of phase a' and that it is 'treated as a fixed function in L∞(Ω)' — no construction is given in §5, and no γ_a values are reported. This is load-bearing because when neighboring structures have nearly identical intensities, (I−c_a)² is nearly equal across phases, so the only term that breaks label symmetry and localizes each phase is γ_a D_a. If D_a is derived from the target anatomy, ground-truth labels, or per-image hand tuning, then the comparison with CV, RSF, CP-ICTM, MBE, and deep learning methods is not equal-footing: the variational model is given spatial information that the baselines do not receive. The claimed 'training-free, competitive with supervised deep learning' result, and specifically the stable HD95 behavior in Table 3, could then reflect the prior rather than CH phase separation. The abstract's 'improved boundary localization' is also internally strained by Table 3, where Ours has mean HD95 3.04 versus U-Net 2.35 and TransUNet 2.55, although Ours does have the smallest SD. The CH-removal ablation (Figure 6) supports the qualitative role of CH regularization, but it does not isolate CH from D_a because D_a is present in both the ablated and complete models.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multiphase variational segmentation model that couples a softmax region-fidelity term with Cahn–Hilliard phase-field regularization, and evolves it by a mixed L2–H^{-1} gradient flow. The authors derive a continuous energy dissipation law, prove existence and uniqueness of weak solutions, and construct a stabilized SAV–FFT scheme with an exact discrete modified-energy identity. Experiments on synthetic, noisy, and SKI10 medical images compare the method with CV, RSF, CP-ICTM, MBE, and several supervised deep-learning baselines, reporting competitive Dice/IoU and the smallest HD95 standard deviation. The main open issue is that the distance-prior term D_a, which is load-bearing when intensity contrast degenerates, is never specified in the experimental section.","tokens_in":22187,"tokens_out":9831,"duration_ms":98182,"significance":"If the mathematical results are correct, the paper provides a sound well-posedness and stability analysis for a nontrivial softmax–Cahn–Hilliard segmentation model, and the SAV–FFT scheme is a useful computational contribution. The empirical claim of stable boundary localization is potentially interesting, but its significance depends on whether the distance-prior D_a is a generic, fairly specified input or a form of leakage of target anatomy. The existence, uniqueness, and discrete-stability proofs are detailed and appear technically correct, which is a genuine strength. The paper does not supply code or reproducible parameter tables, and the current abstract overstates the boundary-localization result relative to the table means.","major_comments":[{"comment":"The distance-prior term γ_a D_a is the only mechanism that breaks label symmetry when (I−c_a)^2 is nearly equal across phases, but the manuscript never specifies how D_a is constructed or what γ_a values are used. The text only says D_a 'encodes spatial prior information' and is treated as a fixed L∞ function. If D_a is built from anatomical knowledge, ground-truth labels, or per-image hand tuning, the comparison with CV/RSF/CP-ICTM and with supervised deep networks in Table 3 is not equal-footing, and the claimed CH-driven weak-boundary separation is not demonstrated. The authors must give the exact construction of D_a, report γ_a, and include an ablation with γ_a=0 (or equivalent) so the role of the CH term is isolated.","section":"§3.1, Eq. (8); §5"},{"comment":"The abstract claims 'improved boundary localization' compared with deep-learning methods, but Table 3 shows mean HD95 = 3.0442 for Ours versus 2.3495 for U-Net and 2.5469 for TransUNet. The data support the more modest claim of the smallest standard deviation (0.4827), not an improvement in mean boundary error. Please rephrase the claim, or provide a paired statistical analysis and a clearly defined notion of 'improvement' (e.g., robustness).","section":"Abstract; §5.5, Table 3"},{"comment":"The experimental setup is not reproducible: no values are reported for λ, γ_a, α, C_0, τ, K, the initialization of U, or the final ε and β used in the SKI10 experiments (Tables 1–2 only report convergence behavior for varying ε and β on a synthetic example). Please provide a complete parameter table and the construction of D_a for every experiment. Without these, the quantitative claims in Table 3 cannot be verified or compared on an equal footing.","section":"§5"}],"minor_comments":[{"comment":"Some typos and stylistic issues: 'Hence,we gain' in the proof of Theorem 3.9; 'Miyoun Jung and coauthors' is informal; the inline definition of W(ua) should be typeset unambiguously.","section":"Throughout"},{"comment":"The qualitative statement that the CH term is 'active' when F_data≈0 would benefit from a quantitative measure of the relative magnitudes of the CH and data forces near the weak boundary.","section":"§2.2"},{"comment":"The caption reports 'Dice = –, IoU = –, HD95 = –' for CV, RSF, and CP-ICTM. Please state explicitly whether these methods failed to produce a final segmentation, and ensure the displayed segmentation images for these methods are visible and labeled.","section":"§5.5, Figure 14"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about D_a is legitimate and should be the primary revision target. The mathematical sections are sound and can be accepted as is; the empirical section needs a full specification of D_a and parameters, plus an ablation without the spatial prior. I see no need for a third round if these are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper before reading it: the mathematical core is honest and mostly checks out, and the experimental comparison is not yet trustworthy because a load-bearing input, the per-phase distance-prior D_a in Eq. (8), is never specified. The stress-test note is right to flag this — it is not a nitpick.\n\nWhat is actually new: the specific model — softmax multiphase fitting with a Cahn–Hilliard term, evolved under a mixed L2–H^{-1} flow with a stabilized SAV-FFT scheme — is not in the cited literature. The continuous analysis is real: the weak formulation matches the PDE, the softmax/mass lemmas are correct, the Galerkin existence argument is standard and contains no evident gap, and the uniqueness argument using the (I−Δ)^{-1} norm is coherent. The discrete modified-energy identity in Theorem 4.1 is algebraically consistent. That is a meaningful theoretical contribution, and the authors deserve credit for not hand-fitting constants to the benchmark and then calling it prediction. The CH-removal ablation and the ε/β parameter studies are also genuinely informative and support the qualitative claim that CH helps preserve weak interfaces.\n\nNow the soft spots, in proportion. The D_a issue is not minor. When neighboring structures have nearly identical intensities, (I−c_a)² degenerates and D_a is the only term that breaks label symmetry. The paper says only that D_a is “a fixed function in L∞” and encodes spatial prior information; its construction, and the values of γ_a, appear nowhere in the experiments. If D_a comes from anatomy knowledge, ground-truth labels, or per-image hand tuning, then “training-free, competitive with supervised deep learning” is not an equal-footing claim, and the boundary-stability result may reflect the prior rather than CH phase separation. This is a fixable but essential omission.\n\nTwo smaller issues: the abstract says “improved boundary localization,” yet Table 3 shows mean HD95 of 3.04 for the method versus 2.35 for U-Net and 2.55 for TransUNet; only the standard deviation is best. That framing should be corrected or qualified. Also, “10 selected images” for testing, with 70/15/10 splits and no description of how the 10 were selected, plus no code/data release, makes the quantitative comparison impossible to reproduce. The CH ablation does not isolate CH from D_a because D_a is present in both the ablated and complete models, though it does show the regularization matters.\n\nWho is this for? Researchers working on variational/phase-field segmentation and energy-stable numerical schemes. They will find the analysis useful and may want to build on it. The practical claim needs verification before it enters any claims about clinical usefulness.\n\nMy recommendation: yes, send it to peer review. It is a serious paper with a real theoretical contribution and an experimentally under-specified claim. The referee should be asked to focus on D_a and γ_a: provide their construction, disclose whether any manual or ground-truth information leaks into them, and rerun or qualify the SKI10 comparison with a fixed, described protocol. If those come out clean, this becomes a solid paper.","headline":"Solid applied-math paper with a genuinely new model and a credible well-posedness proof, but the empirical claims rest on an unspecified distance-prior map and a murky SKI10 test protocol; peer-review it, but require the authors to disclose exactly how D_a is built.","tokens_in":22862,"tokens_out":1590,"would_cite":true,"duration_ms":20757,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68U10","35K55","65M12","65M70"],"pacs":[],"model":"deepseek-v4-flash","headline":"Phase-separation model splits homogeneous regions where image contrast is zero","keywords":["image segmentation","Cahn–Hilliard","phase separation","softmax representation","weak boundary","scalar auxiliary variable","energy stability","medical imaging"],"falsifier":"Run the SKI10 segmentation with the distance prior disabled (γ_a = 0) under identical parameters; if the homogeneous bone regions then merge or the contour leaks, the central claim that Cahn–Hilliard supplies the missing separation when the data force vanishes is not supported. A second check: inspect how D_a is constructed—if it derives from ground-truth masks or per-image tuning, the no-training comparison to supervised deep networks is invalid.","tokens_in":21576,"feed_emoji":"🦴","tokens_out":3875,"duration_ms":37685,"temperature":0.7,"pith_summary":"The paper tries to establish that a Cahn–Hilliard phase-field term, added to a softmax region-fitting energy, can separate adjacent structures with nearly identical intensities and a weak shared boundary, even when the image-driven force is essentially zero near that boundary. The mechanism is phase separation: the double-well potential drives each phase variable toward distinct labels while the gradient penalty keeps a coherent interface, so the contour does not leak or merge when intensity contrast alone would produce no force. The authors support this with a mixed L2–H−1 gradient flow, a proof of existence and uniqueness of weak solutions, and a linear, unconditionally energy-stable SAV–FFT scheme. On 10 SKI10 bone images, the method reaches Dice/IoU 0.9816/0.9639 without any network training and has the most stable HD95 (3.0442 ± 0.4827) among U-Net, TransUNet, Swin-UNet, and MedSegDiff. A training-free variational model that rivals supervised deep networks on boundary precision would matter wherever annotated data are scarce.","feed_headline":"Phase-separation model splits homogeneous regions where contrast is zero","feed_subtitle":"Training-free variational model posts stable HD95 of 3.04 on knee bone images, rivaling supervised U-Net.","key_machinery":"The central object is the Cahn–Hilliard energy per phase, E_CH(u_a) = ∫(ε²/2|∇u_a|² + W(u_a)) dx, with double-well potential W(u_a) = ½ u_a²(u_a−1)². Its variational derivative, −ε²Δu_a + W′(u_a), is the force that keeps driving the phase variables toward 0 or 1 when the softmax data force F_a(U) = λβΦ_a(e_a − ē) degenerates near a weak boundary. The mixed gradient flow ∂_t u_a = −(I−Δ)μ_a couples a direct L2 descent with the fourth-order H−1 regularization of the chemical potential, avoiding the mass conservation of classical Cahn–Hilliard dynamics while preserving higher-order interfacial control. The stabilized SAV splitting with Fourier spectral discretization makes each time step a line","core_discovery":"The central claim is that classical region-fitting forces fail for homogeneous-appearance structures because the Chan–Vese data force factorizes as (c1−c2)(2I−c1−c2), so it vanishes when the fitted mean intensities coincide. The paper argues that the Cahn–Hilliard regularization does not merely smooth the evolving interface but actively promotes phase separation: even when the data force F_data ≈ 0 near the desired boundary, the double-well potential and gradient energy continue to push each phase variable toward 0 or 1, maintaining a coherent diffuse interface and preventing the merging of adjacent homogeneous regions. With the softmax representation and a distance-prior map, the model sepa","pith_inferences":["The distance-prior map D_a is never specified in the paper; if it is constructed from the same annotated labels used to train the deep baselines, the 'without network training' comparison is not equal-footing, and the reported separation could be driven by the prior rather than by Cahn–Hilliard phase separation.","A direct test of the mechanistic claim is to rerun the SKI10 experiments with γ_a = 0; if separation collapses when the distance prior is disabled, then the double-well phase-separation term alone does not carry the claimed effect.","The HD95 stability advantage rests on only 10 test images, so it is suggestive but not evidence of generalizability to other anatomies or imaging modalities.","The same mixed-flow, stabilized-SAV construction could be applied to other fourth-order regularizers with the data term held fixed, which would isolate precisely what the double-well potential adds beyond general higher-order smoothing."],"forward_implications":["If the data-driven force degenerates near a weak boundary, the Cahn–Hilliard term keeps pushing the phase variables toward distinct labels, so adjacent homogeneous structures remain separated instead of merging.","The mixed L2–H−1 flow removes the classical Cahn–Hilliard mass-conservation constraint, allowing region sizes to adapt to the image data rather than being fixed by initialization.","The linear SAV–FFT scheme is shown to remain stable even at time step τ = 10⁴, so the fourth-order model is computationally practical for large time steps.","On the SKI10 tests, the method achieves HD95 = 3.0442 ± 0.4827 without training, with the smallest standard deviation among the compared methods, indicating more consistent boundary localization across test images.","The existence and uniqueness theorem guarantees that the proposed gradient-flow evolution is well posed in the natural solution class C([0,T];L²) ∩ L²(0,T;H²_N)."],"fun_headline_variants":["Phase separation splits same-intensity regions with weak boundaries","Segmentation without intensity contrast: phase separation model","Cahn-Hilliard model preserves weak boundaries in homogeneous areas","Training-free phase-field segmentation beats U-Net on knee bones","When data forces vanish, phase separation saves the boundary"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the distance-prior map D_a and its weights γ_a are chosen without using the target labels or per-image hand-tuning; if D_a already encodes where each phase should be, the claimed separation is driven by the prior, not by Cahn–Hilliard phase separation.","fun_headline_variants_meta":{"raw":{"variants":["Phase separation splits same-intensity regions with weak boundaries","Segmentation without intensity contrast: phase separation model","Cahn-Hilliard model preserves weak boundaries in homogeneous areas","Training-free phase-field segmentation beats U-Net on knee bones","When data forces vanish, phase separation saves the boundary"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000289,"raw_usage":{"total_tokens":1532,"prompt_tokens":751,"completion_tokens":781,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":495,"completion_tokens_details":{"reasoning_tokens":703}},"tokens_in":495,"tokens_out":781,"duration_ms":8335,"temperature":1.0,"reasoning_tokens":703,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T05:58:28.576606+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the SKI10 segmentation with the distance prior disabled (γ_a = 0) under identical parameters; if the homogeneous bone regions then merge or the contour leaks, the central claim that Cahn–Hilliard supplies the missing separation when the data force vanishes is not supported. A second check: inspect how D_a is constructed—if it derives from ground-truth masks or per-image tuning, the no-training comparison to supervised deep networks is invalid.","supporting_citations":[],"review_version":1}