{"id":"69ba2858-ae8a-4e7b-81cd-9915c69c2ef0","arxiv_id":"2502.01988","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A differentiable dMRI simulator combined with a spectral auto-encoder lets gradient descent recover synthetic axon meshes (bending, beading, fanning) from simulated diffusion MRI signals.","lead":"ReMiDi reconstructs 3D axon shapes from diffusion MRI signals by making the MRI simulator differentiable and optimizing a compact latent code of the shape. It is a proof-of-concept that could eventually let doctors see brain tissue microstructure without invasive imaging.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Arbitrary-shape claim is unsupported: the SAE latent is trained on three synthetic deformation families, and the paper's own convergence curves show signal matching alone does not determine geometry.","rationale":"The reader's weakest assumption—latent-space coverage—is the same gate that I find most load-bearing, so I agree with the conditional verdict. My stress-test sharpens it in two ways. First, the experimental protocol only holds out parameter values within the three training families; the auto-encoder has never been asked to represent a shape from a different generative family, so 'arbitrary shapes' is an extrapolation rather than an evaluated capability. Second, the inverse problem's ill-posedness is visible inside the paper: signal loss converges long before spatial error and volume stabilize, so a low signal loss can co-occur with the wrong geometry. That is why a concrete out-of-family reconstruction test that reports both signal loss and Chamfer distance is the right way to adjudicate the concern. If the out-of-family test fails, the appropriate claim is 'interpolation within a learned deformation manifold', not arbitrary-shape reconstruction, and the paper should be revised to scope the abstract and Section 1 accordingly. The code-not-linked and missing-baseline issues are secondary and do not change the verdict.","tokens_in":14192,"tokens_out":6358,"duration_ms":64353,"concrete_test":"Train the SAE on only two of the three deformation families (e.g., bending and beading), then run ReMiDi on a fanning reference mesh with the same 315-vertex connectivity. Record the final signal L2 loss and the modified Chamfer distance. If the signal loss reaches the same magnitude as in-family reconstructions while the Chamfer distance stays near the out-of-family baseline (or is more than twice the in-family error), then signal matching does not imply shape recovery and the arbitrary-shapes claim is unsupported. If both metrics are comparable to in-family results, the latent-coverage concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim of arbitrary-shape reconstruction rests on two unproven premises: (i) every plausible reference mesh lies in (or close to) the image D(z) of the SAE decoder, and (ii) minimizing the signal L2 loss ||F(D(z)) - S_ref|| identifies that z. Neither is established. The SAE is trained only on three synthetic deformation families applied to a single 315-vertex cylinder (Sections 4 and 5.1); the test meshes are held-out parameter values from the same families (e.g., odd vs. even bead counts, even vs. odd fanning angles), so the evaluation demonstrates interpolation within the training manifold, not coverage of arbitrary shapes. A decoder with a 16-d latent cannot emit a mesh with different topology, connectivity, or a geometry outside these families, so such a reference is unreachable by construction. Moreover, the paper's own convergence plots (Fig. 4 vs. Fig. 11/12) show signal loss plateauing near iteration 250 while spatial Chamfer error and volume keep changing until roughly iteration 500, which is direct evidence that low signal loss does not uniquely determine geometry. The fanning results (Section 5.1.3, Table 1) already show poor geometry recovery and an MLP baseline that beats ReMiDi, confirming that signal-based optimization can converge to a mesh that matches the signal but not the reference shape.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents ReMiDi, a PyTorch-based differentiable dMRI simulator that solves the Bloch-Torrey PDE via finite elements and matrix formalism, and uses it to solve a mesh-reconstruction inverse problem. The reconstruction is performed by optimizing a 16-dimensional latent code of a spectral auto-encoder (SAE) so that the simulated signal matches a reference signal S_ref. Experiments on synthetic single-axon meshes with bending/twisting, beading, and fanning deformations show that the optimizer can reduce signal loss and, for some deformation families, reduce spatial error relative to the reference mesh. The paper also compares ReMiDi to an MLP baseline and reports ablation studies on the number of diffusion directions and diffusion times. The core forward simulator is validated against SpinDoctor, and the code is made available.","tokens_in":14477,"tokens_out":4987,"duration_ms":49578,"significance":"The differentiable BTPDE simulator is a useful contribution: it enables gradient-based optimization through a physically grounded dMRI forward model, and the matrix-formalism implementation addresses computational cost. The SAE latent-space optimization is a sensible way to regularize the ill-posed inverse problem. However, the significance is limited by three issues: the inverse validation is a self-consistency check using the same simulator to generate the reference signals; the claimed arbitrary-shape reconstruction is tested only within three narrow synthetic deformation families; and the fanning results are substantially worse than the MLP baseline. These issues do not invalidate the forward simulator or the general optimization framework, but they do mean the paper's central claim, as stated in the abstract, is not supported by the evidence.","major_comments":[{"comment":"The claim that ReMiDi can 'reconstruct microstructures of arbitrary shapes represented by finite-element meshes' is not supported. The SAE is trained on three synthetic deformation families applied to a single 315-vertex cylinder (bending/twisting, beading, fanning), and the test meshes are held-out parameter values from the same families (e.g., odd versus even bead counts, even versus odd fanning angles). Because the decoder D(z) can only produce meshes within the learned manifold, any reference mesh with different topology, connectivity, or deformation type is unreachable by construction. The evaluation therefore demonstrates interpolation within the training manifold, not coverage of arbitrary shapes. The paper should either restrict the claim to the studied deformation families or demonstrate reconstruction of out-of-family geometries.","section":"Abstract, Section 1, Section 4, Section 5.1"},{"comment":"The optimization objective in Algorithm 1, ||F(D(z)) - S_ref||^2, is shown to be insufficient to determine geometry. The signal loss converges around iteration 250 (Fig. 4), while the Chamfer distance and volume continue changing until roughly iterations 400-700 (Figs. 11 and 12). This is direct evidence that low signal loss does not uniquely identify the mesh, which is consistent with the paper's own discussion of ill-posedness. The authors should either incorporate additional regularizers or multi-sequence information that can reduce non-identifiability, or explicitly frame the result as one member of a signal-equivalence class rather than a unique reconstruction.","section":"Section 5, Appendix E.1, Appendix E.2, Algorithm 1"},{"comment":"The fanning experiments contradict the general claim of reconstruction capability. For fanning, ReMiDi achieves a modified Chamfer distance of 7.433 +/- 4.844, which is an order of magnitude worse than the MLP baseline (0.743 +/- 0.24), and Figure 7 shows significant errors at 60 degrees and non-smooth curvature. The text states that 'ReMiDi can capture fanning to some extent,' but this is a much weaker statement than the abstract's claim of reconstructing fanning fibers. Either the method needs to be improved for fanning, or fanning should be explicitly excluded from the headline claims.","section":"Section 5.1.3, Table 1"},{"comment":"The inverse-problem validation is a self-consistency check: the reference signals S_ref are generated by the same differentiable simulator F used for reconstruction. Figure 2 validates the forward simulator against SpinDoctor, but the inverse procedure is not tested against an independent forward model. Thus the experiments show that the optimizer can invert the exact forward model on a training manifold, but do not demonstrate that the method is robust to model error or to signals generated by a different simulator. The authors should validate the inverse approach using an independent forward model or an analytic signal expression, or clearly state this as a limitation.","section":"Section 3.1, Figure 2"}],"minor_comments":[{"comment":"There is a typo: 'unidimentional' should be 'unidimensional'.","section":"Section 1"},{"comment":"The caption reports negative modified Chamfer distances at 60 degrees due to KDE-based smoothing; plotting negative distances for a distance metric is confusing and should be replaced with raw values or an explicit explanation of the smoothing artifact.","section":"Figure 7 caption"},{"comment":"Algorithm 1 hardcodes the input mesh as M in R^{315x3}, while the text uses a general V; this inconsistency should be fixed.","section":"Algorithm 1, Section 4"},{"comment":"The description of the spectral auto-encoder should clarify whether the 300 spectral coefficients are computed per Cartesian coordinate, and how the graph Laplacian and its eigenvectors are defined for the deforming mesh; the appendix says connectivity is consistent but the relationship between geometry changes and the fixed spectral basis should be stated explicitly.","section":"Section 4, Appendix D"},{"comment":"The MLP baseline comparison lacks details on the MLP's input representation, training set, and loss function; without these details, the numbers in Table 1 are difficult to interpret.","section":"Section 5.2"},{"comment":"The main text says the mesh volume converges around iteration 400, while Appendix E.2 says it stabilizes around iteration 500; please reconcile these numbers.","section":"Section 5, Appendix E.2"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline major revision. The forward differentiable simulator and the latent-space optimization idea are sound and potentially useful, but the central 'arbitrary shapes' claim is overstated and the fanning results are poor. The paper would be substantially stronger if the authors reframed it as a proof-of-concept on three deformation families, added an independent forward-model validation, and addressed the non-identifiability issue revealed by the convergence curves. I would not reject the paper because the simulation contribution is solid, but the scope of the claims must be brought in line with the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The differentiable SpinDoctor reimplementation is a real engineering contribution, and the paper earns a serious referee. But the \"arbitrary shapes\" claim in the abstract is a bridge too far given the latent space is trained on three synthetic deformation families of a single 315-vertex cylinder.\n\nWhat's new: a PyTorch implementation of a BTPDE/FEM dMRI simulator, validated against MATLAB SpinDoctor, with matrix formalism and differentiable eigendecomposition via Xi-Torch. That makes gradient-based inversion through the physics possible. Combining it with a spectral auto-encoder to optimize a 16-d latent is the clever part. The experiments are honest: they show signal loss and spatial Chamfer loss over iterations, and they note the mismatch in convergence times (signal plateaus ~250, geometry ~500+). That's a candid admission of ill-posedness.\n\nSoft spots: 1) The \"arbitrary shapes\" claim is unsupported. The SAE is trained on bending, fanning, beading only; test meshes are held-out parameters from the same families. A 16-d latent over a fixed connectivity cannot represent different topology or deformation families, so the method can only interpolate within those learned manifolds. The fanning results (MLP beats ReMiDi; high Chamfer at 60 degrees) confirm the limits. 2) The paper's own figures show that signal matching alone does not pin down geometry; the reconstruction keeps changing geometry long after signal loss is flat. That's an important caveat for anyone reading the abstract as \"we recover the true mesh.\" 3) Reproducibility: the abstract says source code is available, but I found no link in the PDF. That should be fixed.\n\nWhat I'd ask in review: narrow the contribution to the three deformation families (or demonstrate generalization), report reconstruction variability across initializations, and provide the code. The conclusion is actually appropriately cautious (\"proof of concept with synthetic data and simplified voxel structures\"); the abstract and impact statement should be brought in line with it.\n\nBottom line: a solid proof-of-concept for differentiable dMRI-based mesh reconstruction within a constrained shape space. It deserves peer review with major revisions, not a desk reject. I'd cite it if I needed a reference for differentiable dMRI simulation, but I wouldn't rely on the arbitrary-shape claim.","headline":"Useful differentiable dMRI simulator and a clever latent-space inversion, but the 'arbitrary microstructure' claim goes well beyond what the three synthetic deformation families in the latent space can support.","tokens_in":14997,"tokens_out":2394,"would_cite":false,"duration_ms":23719,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ReMiDi reconstructs axonal microstructure meshes from diffusion MRI signals via gradient-based optimization in a learned latent space.","keywords":["diffusion MRI","microstructure reconstruction","Bloch-Torrey PDE","finite element method","spectral auto-encoder","differentiable simulation","axonal geometry","latent space optimization"],"falsifier":"Take a reference mesh that is not a bent, beaded, or fanned cylinder—for example, an axon with an elliptical cross-section, a branch point, or a wavy surface with high-frequency features—generate its dMRI signal with the same simulator, and run ReMiDi from a straight cylinder; if the signal loss reaches near zero while the Chamfer distance to the reference remains large, the claim that ReMiDi can reconstruct arbitrary microstructure shapes is refuted.","tokens_in":14046,"feed_emoji":"🧠","tokens_out":11324,"duration_ms":96548,"temperature":0.7,"pith_summary":"ReMiDi aims to solve the inverse problem of diffusion MRI: given a measured signal, recover the 3D shape of the axons that produced it. The method builds a differentiable forward simulator that computes the diffusion signal from a finite-element mesh, then optimizes the mesh so its simulated signal matches the reference. To avoid the ill-posedness of tweaking every vertex, it encodes the mesh into a low-dimensional latent space using a spectral auto-encoder and runs gradient descent on that latent vector. On synthetic axon meshes with bending, beading, and fanning deformations, the recovered meshes match the reference signal and approach the true geometry, with lower Chamfer distance than a neural network baseline on bending and beading.","feed_headline":"ReMiDi turns dMRI signals back into 3D axon meshes","feed_subtitle":"A differentiable simulator and learned latent space turn diffusion MRI signals into 3D axon shape estimates.","key_machinery":"The load-bearing mechanism is the coupling of a differentiable diffusion MRI simulator with a spectral auto-encoder (SAE). The simulator solves the Bloch-Torrey partial differential equation on a tetrahedral mesh using the matrix formalism—projecting the solution onto a truncated eigenbasis of the Laplace operator—so that the output signal is differentiable with respect to the mesh vertex coordinates. The SAE encodes the 315-vertex mesh into 300 spectral coefficients and then into a 16-dimensional latent vector $\\mathbf{z}$, and its decoder $D$ maps $\\mathbf{z}$ back to a mesh; optimization is performed on $\\mathbf{z}$ with the loss $\\lVert F(D(\\mathbf{z})) - S_{\\mathrm{ref}}\\rVert_2^2$, which confines the search to a learned manifold of plausible axon shapes.","core_discovery":"The paper's central claim is that a diffusion MRI signal contains enough information to recover the geometry of a single axon-like microstructure, provided the search is restricted to a plausible shape manifold. ReMiDi demonstrates this by combining a differentiable finite-element solver of the Bloch-Torrey equation with a spectral auto-encoder: the mesh is mapped to spectral coefficients and then to a 16-dimensional latent vector, and the latent vector is updated by gradient descent on the difference between the simulated and reference signals. The reconstruction converges to the reference signal, and for the deformation families considered—bending and twisting, beading, and fanning—the recovered mesh converges in Chamfer distance to the ground truth, with the most accurate results for beading and the least for fanning at large angles.","pith_inferences":["I interpret the claim of reconstructing 'arbitrary shapes' as an aspiration: the experiments cover only three deformation families on a single 315-vertex cylinder, so the method's generalizability to other topologies is untested. A stronger test would use meshes with heterogeneous cross-sections or branching.","Because the diffusion signal is invariant under certain rotations and reflections—the authors even adjust their evaluation metric accordingly—the objective function may have multiple equivalent optima; the latent prior then determines which orientation or mirror image is recovered, which may or may not match the true mesh.","A practical extension would be to train the SAE on a broader corpus of realistic neuron geometries and then test ReMiDi on signals from those shapes, which would reveal whether the latent space is the bottleneck or whether the simulator is the limiting factor.","The paper's finding that two diffusion times improve reconstruction hints at what information the signal carries about length scales; combining more diffusion times with variable b-values could push the resolution limit further, but the computational cost grows with the number of gradient evaluations."],"forward_implications":["dMRI signals can be inverted into explicit 3D mesh reconstructions of axonal geometry, surfacing features such as undulation, focal beading, and fanning that parametric models only summarize as scalar metrics.","Because the forward simulator is differentiable, the same pipeline can be turned to other gradient-based tasks, such as estimating diffusion coefficients or optimizing acquisition parameters jointly with the mesh.","The success of latent-space optimization in this setting supports the view that learned priors can regularize PDE-constrained inverse problems, making vertex-level reconstruction unnecessary.","The ablation results imply that acquisition design matters: using at least two diffusion times and 15 or more gradient directions substantially improves the spatial fidelity of the recovered mesh.","The comparison with a neural network suggests that physics-based simulation contributes the most when deformations are large; for shapes close to the straight cylinder, a direct regressor can be more accurate."],"supporting_citations":[{"why":"The finite-element dMRI simulator (SpinDoctor) that ReMiDi reimplements in PyTorch and validates against in Figure 2.","marker":"Li et al., 2019"},{"why":"Introduces the matrix formalism that ReMiDi uses to solve the Bloch-Torrey equation semi-analytically for fast simulation.","marker":"Callaghan, 1997"},{"why":"Defines the spectral auto-encoder that ReMiDi uses to map meshes to spectral coefficients and a low-dimensional latent code.","marker":"Lemeunier et al., 2022"},{"why":"Provides the differentiable eigen-decomposition routine used to compute the Laplace eigenvalues and eigenvectors inside the forward pass.","marker":"Kasim & Vinko, 2020"},{"why":"Gives the Bloch-Torrey partial differential equation that governs the diffusion-weighted magnetization evolution simulated by ReMiDi.","marker":"Torrey, 1956"},{"why":"Defines the Chamfer distance used as the spatial reconstruction error metric in all experiments.","marker":"Barrow et al., 1977"},{"why":"Motivates optimizing PDE-constrained inverse problems in a learned latent space, which ReMiDi adapts for mesh reconstruction.","marker":"Zhao et al., 2022"},{"why":"Provides the PyTorch auto-differentiation framework that makes the simulator differentiable and the gradient-based mesh update possible.","marker":"Paszke et al., 2019"}],"fun_headline_variants":["ReMiDi: dMRI signals to 3D axon meshes","Differentiable simulator reconstructs axon geometry","Gradient descent turns dMRI into microstructure shapes","ReMiDi inverts diffusion MRI with learned latent space","From diffusion signals to 3D axon shapes via auto-encoder"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The latent space learned by the spectral auto-encoder is assumed to contain a shape close enough to the true microstructure, and that auto-encoder was trained only on three families of synthetic deformations of a single cylinder, so any real or complex shape outside that manifold cannot be recovered.","fun_headline_variants_meta":{"raw":{"variants":["ReMiDi: dMRI signals to 3D axon meshes","Differentiable simulator reconstructs axon geometry","Gradient descent turns dMRI into microstructure shapes","ReMiDi inverts diffusion MRI with learned latent space","From diffusion signals to 3D axon shapes via auto-encoder"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000184,"raw_usage":{"total_tokens":1322,"prompt_tokens":951,"completion_tokens":371,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":291}},"tokens_in":567,"tokens_out":371,"duration_ms":4181,"temperature":1.0,"reasoning_tokens":291,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:45:46.833103+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a reference mesh that is not a bent, beaded, or fanned cylinder—for example, an axon with an elliptical cross-section, a branch point, or a wavy surface with high-frequency features—generate its dMRI signal with the same simulator, and run ReMiDi from a straight cylinder; if the signal loss reaches near zero while the Chamfer distance to the reference remains large, the claim that ReMiDi can reconstruct arbitrary microstructure shapes is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the Bloch-Torrey partial differential equation that governs the diffusion-weighted magnetization evolution simulated by ReMiDi."},{"cited_title":"G., Tenenbaum, J","cited_arxiv_id":null,"evidence_quote":"Defines the Chamfer distance used as the spatial reconstruction error metric in all experiments."},{"cited_title":"Pytorch: an imperative style, high-performance deep learning library","cited_arxiv_id":null,"evidence_quote":"Provides the PyTorch auto-differentiation framework that makes the simulator differentiable and the gradient-based mesh update possible."}],"review_version":1}