Pith. sign in

REVIEW 3 major objections 6 minor 20 references

Multiscale Parallel Simulation of Malignant Pleural Mesothelioma via Adaptive Domain Partitioning -- an Efficiency Analysis Study

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read A four-core parallel PDE solver accelerates a Cellular Potts model of malignant pleural mesothelioma by 1.8–1.95×, and the dynamic bounding box around the tumour is what lets the lung-scale simulation run on a personal computer.

desk verdict Plausible speedups undercut by missing validation of the dynamic bounding box. read the letter →

arxiv 2505.03067 v1 pith:5KWSC5SP submitted 2025-05-05 cs.CE cs.DCq-bio.QM

classification cs.CEcs.DCq-bio.QM MSC 65Y0592C50
keywords malignantpleuralmesotheliomacellularPottsmodelparallelPDEsolvingmpi4pyFiPydynamicboundingboxspeedupanalysiscomputationaloncology
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims to make a multiscale model of malignant pleural mesothelioma — a Cellular Potts Model of tumour proliferation coupled to reaction-diffusion equations for oxygen, nutrients, and cytokines in the pleural space — computationally tractable on ordinary hardware. The concrete result is a measured speedup of about 1.8 on a $100^3$ mesh and 1.95 on a $200^3$ mesh when the PDE solve switches from serial FiPy to a four-core mpi4py-parallel FiPy run, with four cores the empirical optimum and efficiency falling below 20% at eight or more cores. The paper further asserts that the dynamic bounding box around the growing tumour, and it alone, is what lets the simulation run on a personal computer rather than exhausting the memory of the full pleural-space domain. If correct, the framework turns a cluster-scale workload into a workstation-scale one, opening the door to repeated parametric and patient-geometry studies of MPM.

What carries the argument

Two mechanisms carry the argument. The dynamic bounding box is an axis-aligned 3D box built from the current coordinates of all tumour cells plus an unspecified margin meant to cover diffusion; it is refreshed every 50 Monte Carlo steps and defines the sub-region of the pleural-space mask on which the PDEs are actually solved, shrinking memory and CPU demand as the tumour moves and grows. The parallel solve is the second mechanism: FiPy's internal domain decomposition driven by mpi4py, with a linear GMRES solver over PETSc for convergence, so that each MPI rank reads the shared mesh file, solves its assigned subdomain, and returns updated concentrations to the coordinating process. The reported metrics — parallel speedup $S_p = T_{\text{serial}}/T_{\text{parallel}}$, efficiency $E_p = S_p/p$, and the fractional load imbalance — are what the efficiency analysis actually measures, with the bounding box responsible for the memory reduction and the parallel solver for the measured time reduction.

What would settle it

Run the oxygen, nutrient, and cytokine equations at a fixed tumour state on the full pleural-space mask and inside the dynamic bounding box at the same Monte Carlo step, and compare the fields on the box interior; a relative $\ell^2$ difference of more than a few percent, or any divergence in CPM proliferation and mitosis outcomes over a few hundred steps, would show the bounded problem is not the same problem. A second check is strong scaling on one fixed mesh (for example $200^3$) out to 48 or more cores, to confirm where the speedup curve crosses below one and whether the decline matches the paper's communication-overhead explanation.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is a two-part efficiency result for a coupled CPM-PDE model of mesothelioma growth. First, solving the oxygen, nutrient, and cytokine PDEs with FiPy parallelized via mpi4py shortens per-timestep solving time by a factor of roughly 1.8 on a $100^3$ mesh and 1.95 on a $200^3$ mesh when four cores are used; four cores is the optimum for these meshes, and beyond eight cores the speedup curve declines as communication overhead dominates. Second, the dynamic bounding box — an axis-aligned region computed from the tumour cells' current coordinates plus a diffusion margin, updated every 50 Monte Carlo steps — confines the finite-volume computation to the region of interest, and the paper states that this restriction alone makes the full simulation, with roughly one to eight million cells, feasible on a personal computer. The speedup figures are the headline; the bounding-box claim is the enabler that keeps the memory footprint small enough for the whole framework to run at all.

Load-bearing premise

The load-bearing premise is that the concentration fields computed inside the dynamic bounding box, with its unspecified margin around the tumour cells, match what a full pleural-space solve would produce closely enough that cell behaviour is unchanged; the paper asserts this without a numerical comparison, so if the margin is too small the measured speedup comes at the price of altered biology.

Editorial extensions

If this is right

  • For tumour models in the tested range of about one to eight million cells, switching the PDE solve from serial to a four-core parallel run cuts per-timestep solving time by roughly half, a gain that holds across all three mesh sizes.
  • Four cores is the practical optimum on meshes of $100^3$ to $200^3$; using eight or more cores makes the solve slower, so the framework's rule of thumb is to parallelize modestly rather than maximally.
  • The dynamic bounding box, not the parallel solver, is what opens the door to running repeated simulations on a workstation, which is the precondition for parametric variation studies and patient-specific geometries.
  • Since the $200^3$ mesh shows slightly better speedup and efficiency than the $100^3$ mesh, the parallel benefit should grow as bounding-box domains get larger, as long as memory stays within reach.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A speedup of about 1.9 across four cores is only roughly 50% parallel efficiency, implying that about half of the parallel capacity is lost to partitioning and message passing even at the measured optimum; attributing that loss to FiPy's partitioner versus MPI exchange would show where a better decomposition could buy back time.
  • The bounding box's margin is never given a value, which suggests a natural refinement: size the margin to the diffusion length scale of the slowest chemical species and check the truncated field against the full-domain field, making the restriction's error quantifiable.
  • Because the box refreshes only every 50 Monte Carlo steps, a tumour growing aggressively between refreshes could outrun its margin; coupling the refresh interval or margin to the tumour's measured expansion rate would keep the efficiency gain from silently changing the biology.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper presents an efficiency analysis of a multiscale framework for simulating malignant pleural mesothelioma (MPM) growth, coupling a Cellular Potts Model with reaction-diffusion PDEs for oxygen, nutrients, and cytokines. To make the simulation tractable on a personal computer, the authors introduce a dynamic bounding box around the growing tumour, and they parallelize the PDE solution with FiPy, mpi4py, PETSc, and a GMRES solver. The main quantitative results are that on 4 cores the parallel PDE solver achieves a speedup of about 1.8 for a 100^3 mesh and 1.95 for a 200^3 mesh, while efficiency drops sharply beyond 4 cores. The paper also reports load-imbalance measurements and states that the bounding box alone is what allows the simulation to run on a PC. The central claims are the speedup of the parallel solver and the practical importance of the adaptive bounding box for reducing memory and CPU overhead.

Significance. If the results hold, the framework is a useful engineering contribution for a domain (multiscale tumour simulation) where memory and runtime are known bottlenecks. The paper has clear strengths: it measures three domain sizes and a wide range of core counts, defines explicit metrics (speedup, efficiency, load imbalance), and honestly documents the communication overhead that limits parallelism. The reported speedup, on the other hand, is modest (about 1.8-1.95 on 4 cores), and the more consequential claim about the bounding box is not supported by any numerical validation that the reduced domain preserves the full-domain solution. The absence of repeat measurements and hardware/software details also limits the reproducibility of the quantitative statements. Thus the significance is currently conditional: the efficiency numbers are plausible, but their practical value depends on an unverified correctness condition.

major comments (3)
  1. [Sections 2.3 and 3.5] The central practical claim that the dynamic bounding box lets the simulation run on a personal computer is not supported by any comparison between the restricted-domain solution and the full pleural-space solution. The margin width, the boundary conditions applied at the box faces, and the 50-MCS update interval are parameters that can alter the solutions of Eqs. (1)-(4) inside the tumour region, which in turn changes CPM cell behaviour. The paper does not state the margin size or perform a convergence study in margin width. Please add a numerical experiment that varies the margin and compares the concentration fields and tumour behaviour against a full-domain reference; without this, the reported speedup applies to an unverified reduced problem.
  2. [Sections 3.1-3.3] The quantitative speedup and efficiency curves appear to be based on single measurements with no repetitions, error bars, raw timing data, or statistical analysis, and the hardware/software environment (CPU model and frequency, memory size, operating system, MPI implementation, PETSc version, and solver tolerance) is not reported. For an efficiency analysis study, this is a significant reproducibility gap. Please report the experimental setup in detail, run each configuration multiple times, and state whether the plotted values are means, medians, or single runs.
  3. [Section 2.5, Eq. (7)] The load imbalance definition is not mathematically well-posed as written. Equation (7) states fl,i = (tm_i - (Ti,s/p)) / <tm_i> = tm_i/<ti> - 1, but the equality does not follow from the text's definitions, and the symbols tm_i, Ti,s, <tm_i>, and <ti> are not all defined consistently. Since Figure 6 and the discussion in Section 3.4 rely on this metric, please provide a correct, self-contained definition with consistent notation and define every symbol.
minor comments (6)
  1. [Section 4] The sentence 'with a linear speedup of 2 when using 4 cores' contradicts the measured speedup of 1.8-1.95 reported in Section 5 and Figure 4; please rephrase to avoid implying an ideal speedup that was not observed.
  2. [Equations (1)-(4)] The oxygen concentration is denoted CO in the text but C0 in Eq. (1), with a corresponding S0 consumption term; please make the notation uniform.
  3. [Sections 2.4-2.5] The GMRES solver tolerance and any preconditioner settings are not reported, although these choices affect both solver time and solution accuracy; please state them for reproducibility.
  4. [Section 2.3] The bounding-box margin is described qualitatively as 'additional margin ... to account for diffusion,' but its size is never given; even if the requested convergence study is added, the margin value used in the reported runs should be stated explicitly.
  5. [Title and Section 2] The phrase 'adaptive domain partitioning' is used to describe the bounding-box reduction of the computational domain, but 'partitioning' more commonly refers to dividing the domain across processes for parallel execution; consider clarifying the terminology to avoid confusion with the parallel domain decomposition described in Section 2.5.
  6. [Figures 3-6] The figures would be easier to interpret with explicit units on the time axis and with markers distinguishing measured points from interpolated curves; currently the reader must infer both the time unit and the nature of the plotted lines.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the speedup and bounding-box claims are direct measurements or definitional consequences, not fitted inputs renamed as predictions.

full rationale

The paper's central claims are (1) a parallel FiPy solver reduces PDE solving time with a measured speedup of 1.8–1.95 on 4 cores for meshes of 100^3–200^3, and (2) the dynamic bounding box reduces the computational domain enough to make the simulation feasible on a personal computer. The speedup is defined in Eq. 5 as Sp = T_serial / T_parallel and is reported as a direct timing measurement, not as the output of a fitted model, an imported uniqueness theorem, or a self-citation chain. No parameter is fitted to a subset of data and then 'predicted' back, and no quantity is defined in terms of the quantity it is supposed to derive. The bounding-box statement in Section 3.5 ('Solely due to the bounding box, the simulation can be run on a personal computer') is a direct consequence of solving the same PDEs on a smaller domain; it is a computational-footprint observation, not a circular derivation of a target result. The paper cites earlier work by some of the same authors (e.g., refs. 8, 15, 16), but these citations provide motivational context about tumour-spheroid and biofilm simulations and do not carry the load of the efficiency claim. A legitimate concern, noted also in the reader's take, is that the paper does not numerically validate that concentration fields computed inside the dynamic bounding box match the full pleural-space solution, and it does not state the margin width or the boundary conditions at the box faces. That is a correctness and validation gap about whether the restricted-domain simulation represents the same biological problem, not a circularity in the efficiency analysis: even a perfectly measured speedup would be independent of that issue. Under the hard rules, the absence of a fitted-input-as-prediction step or a load-bearing self-citation chain means the honest finding is no significant circularity, score 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central performance claim rests on an unvalidated bounding box assumption, unspecified solver tolerances, and an unverified equivalence between serial and parallel solves. The biological PDE coefficients are imported from prior work and are not fitted here, so they do not count as free parameters for the efficiency claim. No new entities are introduced.

free parameters (3)
  • bounding box margin
    The additional margin around tumour cell coordinates applied in Section 2.3 controls the size of the PDE domain and therefore the speedup and memory footprint. Its value is not stated, so the reported efficiency numbers are conditional on an unspecified hand-chosen margin.
  • bounding box update interval = 50 MCS
    The dynamic bounding box is recomputed every 50 Monte Carlo steps (Section 2.3). This interval is a hand-chosen trade-off between accuracy and overhead and directly affects the computational time and the validity of the concentration fields.
  • GMRES solver tolerance
    The iterative GMRES solver is mentioned in Section 2.4 without convergence tolerance or preconditioner details, which affects iteration counts and therefore timing measurements.
assumptions (4)
  • domain assumption The reaction-diffusion PDEs (Eqs. 1-4) adequately describe oxygen, nutrient and cytokine dynamics for the simulation's purpose.
    Invoked in Section 2.2; the efficiency study does not validate the biological model against experimental or clinical data, but the performance claim does not depend on biological fidelity.
  • ad hoc to paper The dynamic bounding box around the tumour preserves the full-domain PDE solution because the margin captures all relevant solute gradients.
    This is the load-bearing assumption behind the adaptive partitioning (Sections 2.3 and 3.5). If the margin is too small, the restricted solution diverges from the full pleural-space solution; the paper does not test this.
  • domain assumption The parallel implementation solves the same discrete problem as the serial implementation, so the measured speedup is a fair comparison.
    Section 2.5 describes the parallel workflow but does not show that the serial and parallel solvers produce identical concentration fields or converge to the same tolerance on identical subdomains.
  • standard math Standard numerical methods (finite-volume discretization, implicit Euler, GMRES) converge as expected for this problem.
    The solver setup in Section 2.4 relies on standard methods with no proof of convergence for the specific coupled CPM-PDE problem.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multiscale Parallel Simulation of Malignant Pleural Mesothelioma via Adaptive Domain Partitioning -- an Efficiency Analysis Study." pith.science (2026). https://pith.science/paper/5KWSC5SP

@misc{pith2026250503067,
  author       = {Pith},
  title        = {Pith review of: Multiscale Parallel Simulation of Malignant Pleural Mesothelioma via Adaptive Domain Partitioning -- an Efficiency Analysis Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5KWSC5SP}},
  note         = {Machine review of arXiv:2505.03067}
}
read the original abstract

A novel parallel efficiency analysis on a framework for simulating the growth of Malignant Pleural Mesothelioma (MPM) tumours is presented. Proliferation of MPM tumours in the pleural space is simulated using a Cellular Potts Model (CPM) coupled with partial differential equations (PDEs). Using segmented lung data from CT scans, an environment is set up with artificial tumour data in the pleural space, representing the simulation domain, onto which a dynamic bounding box is applied to restrict computations to the region of interest, dramatically reducing memory and CPU overhead. This adaptive partitioning of the domain enables efficient use of computational resources by reducing the three-dimensional (3D) domain over which the PDEs are to be solved. The PDEs, representing oxygen, nutrients, and cytokines, are solved using the finite-volume method with a first-order implicit Euler scheme. Parallelization is realized using the public Python library mpi4py in combination with LinearGMRESSolver and PETSc for efficient convergence. Performance analyses have shown that parallelization achieves a reduced solving time compared to serial computation. Also, optimizations enable efficient use of available memory and improved load balancing amongst the cores.

Figures

Figures reproduced from arXiv: 2505.03067 by the authors.

Figure 1
Figure 1. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 1
Figure 1. Flowchart depicting the multiscale modelling workflow, starting from organ￾level CT segmentation and progressing through tissue-level domain definition, cell￾level CPM dynamics and parallel continuum-level PDE computations. Everything takes place in the Compucell3D Steppables with exception of the Parallel execution. The lung data are acquired from segmented CT scans using TotalSegmenta￾tor [19,2] and used in the co… view at source ↗
Figure 2
Figure 2. Computational domain generated from CT scans. (A) Zoomed in MPM tumour. (B) Visceral pleura extracted from CT scan segmentation. (C) Tumour in the pleural space. For visualization purposes the parietal pleura is not shown. (D) Bounding box around the tumour, used to reduce the computational demand. 2.2 Mathematical Models For the simulation of tumour growth and cell proliferation we implement a CPM, as visualized in… view at source ↗
Figures from the paper (5 more)
Figure 3
Figure 3. Figure 3: PDE solver computational time per time step for three different domain sizes using different number of cores 3.2 Parallel Speedup In [PITH_FULL_IMAGE:figures/full_fig_p008_3.png]
Figure 4
Figure 4. Figure 4: Parallel speedup for various numbers of cores for the three respective domain sizes. 3.4 Load Imbalance In [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Parallel efficiency for various numbers of cores for the three respective domain sizes. and reduces the number of cores required to run the simulation efficiently, thus reducing the overall computational time [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Load imbalance for various numbers of cores for the three respective domain sizes. and on the number of variables that need to be solved in that domain, as more variables result in more computational work. The bounding box strategy has proved to be of key importance fo…
Figure 7
Figure 7. Figure 7: A contour plot of the normalized nutrient concentrations in the pleural space (a 2D slice is shown), confined by the dynamically set bounding box. The perimeter lines of the tumour cells are highlighted in red. 5 Conclusion We showed that the utilization of a parallel …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 18 canonical work pages

  1. [1]

    PloS one7(3), e33726 (2012)

    Andasari, V., Roper, R.T., Swat, M.H., Chaplain, M.A.: Integrating intracellular dynamics using compucell3d and bionetsolver: applications to multiscale modelling of cancer cell growth and invasion. PloS one7(3), e33726 (2012)

  2. [2]

    The Can- cer Imaging Archive (2015).https://doi.org/10.7937/K9/TCIA.2015.LO9QL9SX, https://doi.org/10.7937/K9/TCIA.2015.LO9QL9SX

    Armato III, S.G., McLennan, G., Bidaut, L., et al.: Data from lidc-idri. The Can- cer Imaging Archive (2015).https://doi.org/10.7937/K9/TCIA.2015.LO9QL9SX, https://doi.org/10.7937/K9/TCIA.2015.LO9QL9SX

  3. [3]

    Computer physics communications 176(11-12), 670–681 (2007)

    Chen, N., Glazier, J.A., Izaguirre, J.A., Alber, M.S.: A parallel implementation of the cellular potts model for simulation of cell-based morphogenesis. Computer physics communications 176(11-12), 670–681 (2007)

  4. [4]

    Computing in Science & Engineering23(4), 47–54 (2021)

    Dalcin, L., Fang, Y.L.L.: mpi4py: Status update after 12 years of development. Computing in Science & Engineering23(4), 47–54 (2021)

  5. [5]

    Annual review of biomedical engineering13(1), 127–155 (2011)

    Deisboeck, T.S., Wang, Z., Macklin, P., Cristini, V.: Multiscale cancer modeling. Annual review of biomedical engineering13(1), 127–155 (2011)

  6. [6]

    Computing in Science & Engineering11(3), 6–15 (2009)

    Guyer, J.E., Wheeler, D., Warren, J.A.: Fipy: Partial differential equations with python. Computing in Science & Engineering11(3), 6–15 (2009)

  7. [7]

    PLOS Computational Biology 20(3), e1011944 (2024)

    Jayathilake, P.G., Victori, P., Pavillet, C.E., Lee, C.H., Voukantsis, D., Miar, A., Arora, A., Harris, A.L., Morten, K.J., Buffa, F.M.: Metabolic symbiosis between oxygenated and hypoxic tumour cells: An agent-based modelling study. PLOS Computational Biology 20(3), e1011944 (2024)

  8. [8]

    Journal of Computational Science54, 101421 (2021)

    Kaura, P., Mishra, T., Verma, N., Dalal, I.S., Sheraton, V.: Effects of combined chemotherapeutic drugs on the growth and survival of cancerous tumours–an in- silico study. Journal of Computational Science54, 101421 (2021)

Show all 20 references
  1. [9]

    Journal of Parallel and Distributed Computing 68(5), 596–608 (2008)

    Korkhov, V.V., Krzhizhanovskaya, V.V., Sloot, P.M.: A grid-based virtual reac- tor: Parallel performance and adaptive load balancing. Journal of Parallel and Distributed Computing 68(5), 596–608 (2008)

  2. [10]

    In: International Conference on Parallel Computing Technologies

    Krzhizhanovskaya, V.V., Korkhov, V.V.: Dynamic load balancing of black-box ap- plications with a resource selection mechanism on heterogeneous resources of the grid. In: International Conference on Parallel Computing Technologies. pp. 245–

  3. [11]

    In: International Con- ference on Parallel Processing and Applied Mathematics

    Krzhizhanovskaya, V.V., Zatevakhin, M.A., Ignatiev, A., Gorbachev, Y.E., Sloot, P.M.: Distributed simulation of silicon-based film growth. In: International Con- ference on Parallel Processing and Applied Mathematics. pp. 879–887. Springer (2001)

  4. [12]

    NPJ Systems Biology and Applica- tions 10(1), 71 (2024)

    Ma, C., Gurkan-Cavusoglu, E.: A comprehensive review of computational cell cycle models in guiding cancer treatment strategies. NPJ Systems Biology and Applica- tions 10(1), 71 (2024)

  5. [13]

    Journal of mathematical biology58, 765–798 (2009)

    Macklin, P., McDougall, S., Anderson, A.R., Chaplain, M.A., Cristini, V., Lowen- grub, J.: Multiscale modelling and nonlinear simulation of vascular tumour growth. Journal of mathematical biology58, 765–798 (2009)

  6. [14]

    SIAM Journal on scientific and statistical computing 7(3), 856–869 (1986)

    Saad, Y., Schultz, M.H.: Gmres: A generalized minimal residual algorithm for solving nonsymmetric linear systems. SIAM Journal on scientific and statistical computing 7(3), 856–869 (1986)

  7. [15]

    BMC cancer20, 1–16 (2020) 14 A

    Sheraton, M.V., Chiew, G.G.Y., Melnikov, V., Tan, E., Luo, K.Q., Verma, N., Sloot, P.: Emergence of spatio-temporal variations in chemotherapeutic drug ef- ficacy: in-vitro and in-silico 3d tumour spheroid studies. BMC cancer20, 1–16 (2020) 14 A. Dolganov et al

  8. [16]

    In: Computational Science–ICCS 2018: 18th International Confer- ence, Wuxi, China, June 11–13, 2018, Proceedings, Part I 18

    Sheraton, M.V., Sloot, P.M.: Parallel performance analysis of bacterial biofilm sim- ulation models. In: Computational Science–ICCS 2018: 18th International Confer- ence, Wuxi, China, June 11–13, 2018, Proceedings, Part I 18. pp. 496–505. Springer (2018)

  9. [17]

    Computing in Science & Engineering 24(5), 53–63 (2022)

    Stack, M., Macklin, P., Searles, R., Chandrasekaran, S.: Openacc acceleration of an agent-based biological simulation framework. Computing in Science & Engineering 24(5), 53–63 (2022)

  10. [18]

    In: Methods in cell biology, vol

    Swat, M.H., Thomas, G.L., Belmonte, J.M., Shirinifard, A., Hmeljak, D., Glazier, J.A.: Multi-scale modeling of tissues using compucell3d. In: Methods in cell biology, vol. 110, pp. 325–366. Elsevier (2012)

  11. [19]

    Radiology: Artificial Intelligence 5(5), e230024 (2023)

    Wasserthal, J., Breit, H.C., Meyer, M.T., Pradella, M., Hinck, D., Sauter, A.W., Heye, T., Boll, D.T., Cyriac, J., Yang, S., et al.: Totalsegmentator: robust segmen- tation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence 5(5), e230024 (2023)

  12. [20]

    Journal of biomechanical engineering 136(2), 021011 (2014)

    Wu, M., Swartz, M.A.: Modeling tumor microenvironments in vitro. Journal of biomechanical engineering 136(2), 021011 (2014)

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.