Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Learned modality gates in multi-modal prostate MRI do not always route by sample quality; their behavior depends on the backbone architecture.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-12 22:23 UTC pith:WKLKVHW5

load-bearing objection Solid empirical mechanism study: gating is backbone-conditional on the two tested stacks, with clean gate-SD and mean-gate counterfactuals; the class-level story is the soft edge, not the measurements. the 3 major comments →

arxiv 2604.10702 v4 pith:WKLKVHW5 submitted 2026-04-12 cs.CV cs.AI

Backbone-Conditional Behavior of Modality Gating in Multi-Modal Prostate MRI Segmentation: A 5-Fold Cross-Validation and Gate Mechanism Analysis

Yongbo Shu , Wenzhao Xie , Shanhu Yao , Zirui Xin , Luo Lei , Kewen Chen , Aijing Luo This is my paper
classification cs.CV cs.AI
keywords prostate cancermulti-modal fusiongated fusionmodality dropoutMRI segmentationcross-validationbackbone-conditional gating
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Multi-modal MRI models for clinically significant prostate cancer often add learned gates that are assumed to weight each sequence according to its quality on that particular scan. This paper tests that assumption directly across two very different backbones—a convolutional network and a state-space (Mamba) network—on a large public dataset with 180 trained models, then checks the gates themselves and their effect under domain shift. On the convolutional backbone the gates collapse into nearly fixed modality weights, carry almost no case-by-case information, and actually lower detection ranking scores. On the Mamba backbone the same gates keep varying from patient to patient, removing that variation hurts performance, and gating plus modality dropout improves scores. Only random modality dropout during training helped both architectures, and the gated Mamba model best retained the ability to correctly leave normal cases alone on an external hospital dataset. The practical message is that gate modules are not a universal quality router; their real behavior is set by the backbone’s built-in way of handling channels.

Core claim

Learned modality gates do not universally implement per-sample quality routing; their effective behavior is backbone-conditional. On nnU-Net the gates collapse into a near-static modality prior (across-case SD 0.0033) whose replacement by the training-set mean leaves performance unchanged and whose addition reduces ranking score (marginal effect −0.037). On Mamba the gates retain sample-dependent variation roughly 11× larger, mean-gate replacement degrades Dice and ranking score, and the gating-plus-dropout configuration improves ranking score by +0.024.

What carries the argument

Modality-Isolated Gated Fusion (MIGF) plus gate-variance and counterfactual mean-gate analysis. MIGF encodes each MRI sequence in an independent bias-free stream so a missing input produces zero features, then learns convex modality weights; measuring the across-case standard deviation of those weights and swapping them for their training-set mean distinguishes a static prior from true per-sample routing.

Load-bearing premise

The result is assumed to reflect the architecture classes themselves, not just the two specific backbones and the single entry-stage scalar gate that were measured.

What would settle it

Apply the identical gated fusion module to several other convolutional and state-space segmentation backbones, compute across-case gate-weight standard deviation, and re-run the mean-gate replacement test; if any convolutional model shows high gate variability and clear performance loss when gates are replaced by their mean, the backbone-class claim is false.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Gate-weight inspection and mean-gate counterfactuals should be routine diagnostics whenever a gated fusion module is added, because the routing interpretation is architecture-contingent.
  • Training-time modality dropout is the only component shown to raise ranking score on both tested backbones and is therefore the architecture-agnostic robustness default for variable diffusion acquisitions.
  • Under cross-cohort shift, convolutional models collapse to near-zero case-level specificity while Mamba models retain it; MIGF-Mamba is the most robust configuration among those tested.
  • Clinical backbone choice must trade that retained specificity against the roughly twelve-fold higher inference latency of the Mamba variant.
  • Explicit gating can be redundant or harmful on convolutional architectures that already reweight input channels through per-channel filters and normalization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Published gains attributed to “adaptive routing” in other gated multi-modal systems built on convolutional U-Nets may sometimes be static priors in disguise; re-inspecting those gates would be a low-cost check.
  • If the inductive-bias account is right, deliberately adding or removing channel-selection bias inside a backbone should predictably flip whether its gates stay static or become sample-dependent.
  • Models whose gates demonstrably react to degraded diffusion scans could supply an auditable quality signal that clinical pipelines could log alongside the segmentation itself.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript reports a dual-backbone factorial study of modality-isolated gated fusion (MIGF) for clinically significant prostate cancer segmentation on PI-CAI (n=1500), with external evaluation on Prostate158 (n=158). Across eight configurations of gating, modality dropout, and deep supervision (180 trained models under 5-fold CV) and a gate-weight/counterfactual analysis of 30 A2 models, the authors find that learned modality gates are backbone-conditional: on nnU-Net they collapse to a near-static modality prior (across-case SD ≈0.0033) and reduce Ranking Score (marginal effect −0.037; A2/A3/A4 p<0.05), whereas on a Mamba backbone they retain sample-dependent variation (~11× larger SD, non-overlapping) and the G+M configuration improves Ranking Score (+0.024, p=0.037). Mean-gate replacement leaves nnU-Net unchanged but degrades Mamba. Modality dropout is the only component beneficial on both backbones; under cross-cohort shift, convolutional models lose case-level specificity while MIGF-Mamba retains the highest among tested configurations.

Significance. If the reported pattern holds, the paper supplies a concrete, falsifiable challenge to a common premise in gated multi-modal medical fusion—that learned gates implement per-sample quality routing—and shows that this premise is architecture-contingent rather than universal. Strengths include the scale of the matched 5-fold factorial (180 models), direct inspection of gate weights with non-overlapping SD distributions, consistent mean-gate counterfactuals (15/15 Mamba models degrade), replication of convolutional specificity collapse on a second U-Net, public code, and an honest limitations section. The architecture-agnostic benefit of training-time modality dropout is immediately actionable for protocols with fragile diffusion sequences. The contribution is complementary to recent prostate-MRI segmentation work and is of clear interest to multi-modal medical imaging and clinical deployment design.

major comments (3)
  1. Discussion §4.1 states that the ≈11× separation in gate variability “indicates that whether a learned gate performs per-sample routing or collapses to a static prior is determined by the inductive bias of the underlying architecture—not by the gating module in isolation,” and attributes this to 3D convolutions already encoding per-channel weighting versus state-space kernels lacking it. The empirical pattern is well supported for the two fully analyzed backbones (nnU-Net and one Mamba instantiation) and the A2 entry-stage scalar gates (§2.4, Table 2, Fig. 3). The class-level mechanistic claim, however, rests on those two instances; the MONAI 3D U-Net was used only for cross-cohort specificity (Table 3), not for gate-SD or counterfactual analysis. Please either (i) temper the language to “the tested convolutional vs state-space instantiations” throughout Abstract/Discussion/Conclusion, or
  2. Section 2.2 defines AdaptiveModalGating as a two-stage module: (1) softmax modality weights α and (2) a feature-level 1×1×1 gate with SiLU/sigmoid. The entire gate-mechanism analysis (§2.4, Table 2, Fig. 3) and the counterfactual mean-gate replacement target only the entry-stage α. The paper’s central claim is about “learned modality gates,” yet the second-stage feature gate is never inspected for sample dependence or subjected to a mean-replacement test. If the feature-level gate carries residual sample-dependent routing on nnU-Net, the “static prior” conclusion would be incomplete. Please either analyze the feature-level gate outputs analogously, or explicitly restrict the claim to the modality-weight stage α and justify why that stage alone is the right object for testing the routing premise.
  3. Table 3 and §3.3 report near-zero Prostate158 case-level specificity for both convolutional families (nnU-Net and MONAI U-Net) versus retained specificity for Mamba (MIGF-Mamba 0.314). CaseSpec is a safety-relevant claim in the Discussion (§4.2), but the manuscript does not state the decision threshold, operating-point selection, or how picai_eval case-level specificity is computed for this external set (e.g., fixed PI-CAI threshold vs re-calibrated). Without that detail, the architecture-conditional “specificity collapse” could partly reflect threshold mismatch under domain shift rather than pure architectural failure. Please specify the CaseSpec definition and threshold protocol for the cross-cohort evaluation, and if a fixed threshold was used, report a brief sensitivity check (e.g., threshold sweep or precision–recall operating points) so the collapse claim is not confounded by calib
minor comments (6)
  1. Several concatenated words appear in the Highlights and early text (e.g., “Counterfactualmean-gatereplacementleavesnnU-Netperformanceun-changed,” “gate-weightand,” “clinicallysignificantprostatecancer”). Please run a full pass for spacing and hyphenation.
  2. Figure 1 has a duplicated caption block (“Figure 1: Overview of the MIGF framework” appears twice with slightly different text). Consolidate into a single caption.
  3. The Introduction and Abstract refer to multi-parametric MRI, while Methods correctly describe PI-CAI as biparametric (T2W, HBV, ADC). Align terminology (biparametric vs multi-parametric) for accuracy.
  4. Table 1: for the six single-seed Mamba configurations, the caption already notes underpowered Wilcoxon tests; consider adding a footnote that p-values are omitted by design so readers do not misread “n.s.” as a null finding.
  5. Fig. 4 is discussed in §4.2 but is not introduced in Results; a one-sentence pointer in §3.1 would help readers locate the degradation scenarios earlier.
  6. Related-work citations for gated multi-modal fusion (mmFormer, CMAF-Net, RFNet) are appropriate; a brief note on whether those works inspected gate outputs would strengthen the “rarely tested directly” claim without expanding scope.

Circularity Check

0 steps flagged

No significant circularity: empirical ablation and gate measurements on held-out folds, not results forced by definition or self-citation.

full rationale

This is a standard empirical multi-modal segmentation study. The central claims (backbone-conditional gate behavior; nnU-Net gates as near-static prior with across-case SD 0.0033 vs Mamba 0.0365; counterfactual mean-gate replacement leaving nnU-Net unchanged while degrading Mamba; marginal ranking-score effects of gating/dropout) are obtained by training 180 models under 5-fold CV, extracting entry-stage gate weights α on validation cases, computing their across-case SD, and re-running inference with training-set mean gates. These are operational measurements and controlled ablations, not quantities that equal their inputs by construction. The isolation property of bias-free streams is an intentional design choice, not a claimed derivation. Marginal component effects are simple averages over the 2^3 factorial, not fitted free parameters re-labeled as predictions. References are external (nnU-Net, Mamba, PI-CAI, Prostate158, ModDrop, etc.); there is no load-bearing self-citation of a uniqueness theorem or ansatz that forces the result. Cross-cohort numbers are pure inference on an external set never used for training. No equation or statistical step reduces the ranking-score or gate-variability claims to a tautology. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central claim rests on standard supervised segmentation practice plus a small set of design choices (MIGF isolation, gate definition, ModDrop rate, two backbone classes). No new physical entities. Free parameters are training hyperparameters and module knobs that affect absolute scores but are shared across the factorial contrasts that drive the backbone-conditional claim. Domain assumptions about mpMRI sequences and evaluation metrics are inherited from PI-CAI/PI-RADS practice.

free parameters (5)
  • ModDrop probability = 0.3
    Stochastic modality zeroing probability set to 0.3 during training; affects robustness ablations and marginal M effect.
  • AdamW learning rate = 5e-5
    Fixed training hyperparameter from picai_baseline recipe; not claimed as universal optimum.
  • DiceFocal loss (alpha, gamma) = alpha=0.9, gamma=2.0
    Class-imbalance weighting chosen by authors; influences absolute Dice/ranking but shared across configs.
  • Deep supervision auxiliary loss weights = 0.5, 0.25
    Auxiliary heads weighted 0.5 and 0.25 when D is enabled; design choice in factorial.
  • Training epochs / no early stopping = 300 epochs
    Fixed 300-epoch schedule with best-validation Ranking Score checkpointing.
axioms (5)
  • ad hoc to paper Bias-free isolated modality encoders yield identically zero features for zero-filled missing inputs (isolation property).
    Core MIGF design claim in §2.2; enables graceful missing-modality behavior by construction.
  • ad hoc to paper Entry-stage softmax modality weights α are the right object for testing per-sample quality routing.
    Gate-mechanism analysis §2.4 targets α; feature-level sigmoid gate is not the primary object of the SD/counterfactual tests.
  • domain assumption PI-CAI Ranking Score (mean of lesion-level AUROC and AP) plus case-level specificity are appropriate clinical proxies for csPCa detection quality.
    Evaluation via picai_eval §2.5; standard in this challenge ecosystem but not the only possible clinical utility metric.
  • domain assumption nnU-Net and the chosen Mamba backbone are representative of convolutional vs state-space classes for modality handling.
    Discussion §4.1 attributes opposite gate behavior to architectural inductive bias; only one model per class in the factorial (plus MONAI U-Net for cross-cohort only).
  • domain assumption Patient-level 5-fold splits and Prostate158 as external cohort adequately probe generalization under domain shift.
    §2.1 and §3.3; single external dataset acknowledged as a limitation.
invented entities (1)
  • Modality-Isolated Gated Fusion (MIGF) / AdaptiveModalGating no independent evidence
    purpose: Decouple per-modality encoding from fusion and provide learnable modality weights plus feature-level gating for multi-parametric MRI.
    Named architecture introduced in §2.2; built from standard conv/MLP/softmax pieces but packaged as the study vehicle. Independent evidence is limited to the paper’s own ablations and public code, not external replications yet.

pith-pipeline@v1.1.0-grok45 · 19249 in / 3673 out tokens · 40748 ms · 2026-07-12T22:23:16.007120+00:00 · methodology

0 comments
read the original abstract

Robust segmentation of clinically significant prostate cancer (csPCa) on multi-parametric MRI must tolerate frequent degradation of its most informative diffusion sequences. Multi-modal fusion commonly employs learned modality gating under the assumption that gates implement per-sample modality quality routing -- rarely tested directly. We ask how gating behaves across backbone architectures. We systematically analyze modality-isolated gated fusion (MIGF) for csPCa segmentation on two backbones (nnU-Net and Mamba) using PI-CAI (n=1500), with cross-cohort validation on Prostate158 (n=158): a factorial ablation over gating, modality dropout, and deep supervision under 5-fold cross-validation (180 trained models), plus a gate-weight and counterfactual analysis of 30 trained gating models. Modality gating is backbone-conditional. On nnU-Net, adding gating reduces the ranking score (marginal effect -0.037; gating configurations p<0.05), whereas on Mamba the gating-plus-dropout configuration improves it (+0.024, p=0.037). Gate-weight analysis explains this: nnU-Net gates collapse into a near-static modality prior (across-case SD 0.0033), while Mamba gates retain sample-dependent variation (0.0365, ~11x larger, non-overlapping); replacing per-sample gates with their training-set mean leaves nnU-Net unchanged but degrades Mamba. Modality dropout is the only component beneficial on both backbones. Under cross-cohort shift, convolutional backbones collapse to case-level specificity near zero, whereas Mamba retains it (MIGF-Mamba highest, 0.31). Learned modality gates do not universally perform per-sample quality routing; their effective behavior is conditional on the backbone's inherent modality awareness. Among tested configurations, MIGF-Mamba is the most cross-cohort robust, and training-time modality dropout is the only component beneficial across both backbones.

Figures

Figures reproduced from arXiv: 2604.10702 by Aijing Luo, Kewen Chen, Luo Lei, Shanhu Yao, Wenzhao Xie, Yongbo Shu, Zirui Xin.

Figure 1
Figure 1. Figure 1: Overview of the proposed Modality-Isolated Gated Fusion (MIGF) framework. [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Detailed structure of the MIGF module. Given modality-specific feature maps [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Robustness profiles of bare and MIGF-equipped backbones across seven evalu [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Demystifying Modality Robustness — The Interplay of Feature Isolation and Modality Dropout [PITH_FULL_IMAGE:figures/full_fig_p034_4.png] view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. A multi-architecture study of specificity refinement and false-positive mechanism analysis in prostate MRI

    eess.IV 2026-06 unverdicted novelty 4.0

    False positives in prostate MRI AI detection share contrast features with true lesions across five architectures, and a lightweight refinement head improves specificity conditionally but not consistently on external data.

Reference graph

Works this paper leans on

28 extracted references · 4 canonical work pages · cited by 1 Pith paper

  1. [1]

    F. Bray, M. Laversanne, H. Sung, J. Ferlay, R. L. Siegel, I. Soerjo- mataram, A. Jemal, Global cancer statistics 2022: GLOBOCAN es- timates of incidence and mortality worldwide for 36 cancers in 185 countries, CA: A Cancer Journal for Clinicians 74 (3) (2024) 229–263. doi:10.3322/caac.21834. 16

    doi:10.3322/caac.21834 2022

  2. [2]

    S. Loeb, A. Vellekoop, H. U. Ahmed, J. Catto, M. Emberton, R. Nam, D. J. Rosario, V. Scattoni, Y. Lotan, Systematic review of complications of prostate biopsy, European Urology 64 (6) (2013) 876–892.doi:10. 1016/j.eururo.2013.05.049

    2013

  3. [3]

    M. A. Bjurlin, J. S. Wysock, S. S. Taneja, Optimization of prostate biopsy: Review of technique and complications, Urol. Clin. North Am. 41 (2) (2014) 299–313.doi:10.1016/j.ucl.2014.01.011

    doi:10.1016/j.ucl.2014.01.011 2014

  4. [4]

    B.Turkbey, A.B.Rosenkrantz, M.A.Haider, A.R.Padhani, G.Villeirs, K. J. Macura, C. M. Tempany, P. L. Choyke, F. Cornud, D. J. Margolis, H. C. Thoeny, S. Verma, Prostate imaging reporting and data system version 2.1: 2019 update of prostate imaging reporting and data system version 2, Eur. Urol. 76 (3) (2019) 340–351.doi:10.1016/j.eururo. 2019.02.033

    doi:10.1016/j.eururo 2019

  5. [5]

    H. U. Ahmed, A. El-Shater Bosaily, L. C. Brown, R. Gabe, R. Kaplan, M. K. Parmar, Y. Collaco-Moraes, K. Ward, R. G. Hindley, A. Free- man, A. P. Kirkham, R. Oldroyd, C. Parker, M. Emberton, Diagnostic accuracy of multi-parametric mri and trus biopsy in prostate cancer (promis): A paired validating confirmatory study, Lancet 389 (10071) (2017) 815–822.doi:...

    doi:10.1016/s0140-6736(16)32401-1 2017

  6. [6]

    Kasivisvanathan, A

    V. Kasivisvanathan, A. S. Rannikko, M. Borghi, V. Panebianco, L. A. Mynderse, M. H. Vaarala, A. Briganti, L. Budaus, G. Hellawell, R. G. Hindley, M. J. Roobol, S. Eggener, M. Ghei, A. Villers, F. Bladou, P. Jichlinski, L. Klotz, M. Kriegmair, D. E. Neal, et al., Mri-targeted or standard biopsy for prostate-cancer diagnosis, N. Engl. J. Med. 378 (19) (2018...

    doi:10.1056/nejmoa1801993 2018

  7. [7]

    Giganti, V

    F. Giganti, V. Kasivisvanathan, A. Kirkham, S. Punwani, M. Emberton, C. M. Moore, C. Allen, Prostate mri quality: A critical review of the last 5 years and the role of the pi-qual score, Br. J. Radiol. 95 (1131) (2022) 20210415.doi:10.1259/bjr.20210415

    doi:10.1259/bjr.20210415 2022

  8. [8]

    Plodeck, C

    V. Plodeck, C. G. Radosa, H.-M. Hubner, C. Baldus, A. Borkowetz, C. Thomas, J.-P. Kuhn, M. Laniado, R.-T. Hoffmann, I. Platzek, Rectal gas-induced susceptibility artefacts on prostate diffusion-weighted mri with epi read-out at 3.0 t: Does a preparatory micro-enema improve 17 image quality?, Abdom. Radiol. (NY) 45 (2020) 4244–4251.doi:10. 1007/s00261-020-02600-9

    2020

  9. [9]

    A. M. Hötker, H. A. Vargas, O. F. Donati, Abbreviated mr proto- cols in prostate mri, Life (Basel) 12 (4) (2022) 552.doi:10.3390/ life12040552

    2022

  10. [10]

    Brizmohun Appayya, J

    M. Brizmohun Appayya, J. Adshead, H. U. Ahmed, C. Allen, A. Bain- bridge, T. Barrett, F. Giganti, J. Graham, P. Haslam, E. W. John- ston, C. Kastner, A. P. S. Kirkham, A. Lipton, A. McNeill, L. Moniz, C. M. Moore, G. Nabi, A. R. Padhani, C. Parker, A. Patel, et al., Na- tional implementation of multi-parametric magnetic resonance imaging for prostate canc...

    doi:10.1111/bju.14361 2018

  11. [11]

    Zhang, N

    Y. Zhang, N. He, J. Yang, Y. Li, D. Wei, Y. Huang, Y. Zhang, Z.He, Y.Zheng, mmformer: Multimodalmedicaltransformer forincom- plete multimodal learning of brain tumor segmentation, in: L. Wang, Q. Dou, P. T. Fletcher, S. Speidel, S. Li (Eds.), Medical Image Computing and Computer Assisted Intervention – MICCAI 2022, Springer Nature Switzerland, Cham, 2022,...

    2022

  12. [12]

    K. Sun, J. Ding, Q. Li, W. Chen, H. Zhang, J. Sun, Z. Jiao, X. Ni, Cmaf-net: A cross-modal attention fusion-based deep neural network for incomplete multi-modal brain tumor segmentation, Quant. Imaging Med. Surg. 14 (7) (2024) 4579–4604.doi:10.21037/qims-24-9

    doi:10.21037/qims-24-9 2024

  13. [13]

    Y. Ding, X. Yu, Y. Yang, Rfnet: Region-aware fusion network for in- complete multi-modal brain tumor segmentation, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 3975–3984

    2021

  14. [14]

    Isensee, P

    F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, K. H. Maier-Hein, nnu-net: A self-configuring method for deep learning-based biomedical image segmentation, Nat. Methods 18 (2) (2021) 203–211.doi:10. 1038/s41592-020-01008-z

    2021

  15. [15]

    A. Gu, T. Dao, Mamba: Linear-time sequence modeling with selective state spaces, in: First Conference on Language Modeling (COLM), 2024. URLhttps://openreview.net/forum?id=tEYskw1VY2 18

    2024

  16. [16]

    A. Saha, J. S. Bosma, J. J. Twilt, B. van Ginneken, A. Bjartell, A. R. Padhani, D. Bonekamp, G. Villeirs, G. Salomon, G. Giannar- ini, J. Kalpathy-Cramer, J. Barentsz, K. H. Maier-Hein, M. Rusu, O. Rouviere, R. van den Bergh, V. Panebianco, V. Kasivisvanathan, N. A. Obuchowski, D. Yakar, M. Elschot, J. Veltman, J. J. Futterer, M. de Rooij, H. Huisman, Art...

    doi:10.1016/s1470-2045(24)00220-1 2024

  17. [17]

    L. C. Adams, M. R. Makowski, G. Engel, M. Rattunde, F. Busch, P. As- bach, B. Hamm, K. K. Bressem, Prostate158 – an expert-annotated 3t mri dataset and algorithm for prostate cancer detection, Comput. Biol. Med. 148 (2022) 105817.doi:10.1016/j.compbiomed.2022.105817

    doi:10.1016/j.compbiomed.2022.105817 2022

  18. [18]

    K. Zhao, A. Ling Yu Hung, K. Pang, P. Hajipour, H. Wu, S. Raman, K. Sung, PCa-Mamba: Spatiotemporal state space models for prostate cancer detection in multi-parametric MRI, Medical Image Analysis 111 (2026) 104033.doi:10.1016/j.media.2026.104033

    work page doi:10.1016/j.media.2026.104033 2026

  19. [19]

    J. A. Alzate-Grisales, A. Mora-Rubio, M. Perán-Teruel, A. N. Bel- trán, C. R. Torres, J. M. O. García, F. García-García, R. Tabares- Soto, J. M. García-Gómez, M. de la Iglesia-Vayá, Clinically significant prostate cancer detection with deep learning in a multi-center mag- netic resonance imaging study, Scientific Reports 16 (1) (2026) 10976. doi:10.1038/s...

    work page doi:10.1038/s41598-026-42214-7 2026

  20. [20]

    Y. Li, M. Huang, Y. Wu, Y. Zhang, Automatic segmentation and PI- RADS grading of prostate cancer for biparametric MRI, Frontiers in Medicine 13 (2026) 1755311.doi:10.3389/fmed.2026.1755311

    work page doi:10.3389/fmed.2026.1755311 2026

  21. [21]

    E. H. P. Pooch, G. Agrotis, L. Cai, M. Emberton, T. T. Shah, H. U. Ahmed, R. G. H. Beets-Tan, S. Benson, T. Janssen, I. G. Schoots, Semi-supervised learning in prostate mri tumor detection approaches fully supervised performance on external validation, Eur. Radiol. (2026). doi:10.1007/s00330-026-12324-x

    doi:10.1007/s00330-026-12324-x 2026

  22. [22]

    Neverova, C

    N. Neverova, C. Wolf, G. Taylor, F. Nebout, Moddrop: Adaptive multi- modal gesture recognition, IEEE Trans. Pattern Anal. Mach. Intell. 19 38 (8) (2016) 1692–1706.doi:10.1109/TPAMI.2015.2461544. URLhttps://doi.org/10.1109/TPAMI.2015.2461544

    doi:10.1109/tpami.2015.2461544 2016

  23. [23]

    M. J. Cardoso, W. Li, R. Brown, N. Ma, E. Kerfoot, Y. Wang, B. Murrey, A. Myronenko, C. Zhao, D. Yang, V. Nath, Y. He, Z. Xu, A. Hatamizadeh, A. Myronenko, W. Zhu, Y. Liu, M. Zheng, Y. Tang, I. Yang, M. Zephyr, B. Hashemian, S. Alle, M. Z. Darestani, C. Budd, M. Modat, T. Vercauteren, G. Wang, Y. Li, Y. Hu, Y. Fu, B. Gor- man, H. Johnson, B. Genereaux, B....

    Pith/arXiv arXiv 2022

  24. [24]

    Cicek, A

    O. Cicek, A. Abdulkadir, S. S. Lienkamp, T. Brox, O. Ronneberger, 3d u-net: Learning dense volumetric segmentation from sparse annotation, in: Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016, 2016, pp. 424–432.doi:10.1007/978-3-319-46723-8_ 49

    work page doi:10.1007/978-3-319-46723-8_ 2016

  25. [25]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization (2019). arXiv:1711.05101. URLhttps://arxiv.org/abs/1711.05101

    Pith/arXiv arXiv 2019

  26. [26]

    Milletari, N

    F. Milletari, N. Navab, S.-A. Ahmadi, V-net: Fully convolutional neural networks for volumetric medical image segmentation (2016).arXiv: 1606.04797. URLhttps://arxiv.org/abs/1606.04797

    Pith/arXiv arXiv 2016

  27. [27]

    T.-Y. Lin, P. Goyal, R. Girshick, K. He, P. Dollar, Focal loss for dense object detection, in: Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 2980–2988.doi:10.1109/ICCV.2017. 324

    doi:10.1109/iccv.2017 2017

  28. [28]

    DIAGNijmegen, Evaluation utilities for 3d detection and diagno- sis in medical imaging [www document],https://github.com/ DIAGNijmegen/picai_eval, accessed: 2026-03-30 (2022). 20

    2026