Pith. sign in

REVIEW 3 major objections 5 minor 55 references

Modeling how liver lesions enhance and wash out across CT phases, rather than just combining phase images, yields measurably sharper boundaries in automated segmentation.

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

2026-08-03 04:23 UTC pith:7V2H7NCC

load-bearing objection Plausible multi-phase fusion architecture with a thorough ablation, but the headline numbers don't reconcile across tables — resolve that before trusting the gains. the 3 major comments →

arxiv 2607.29568 v1 pith:7V2H7NCC submitted 2026-07-31 cs.CV

DynoDINO: Harnessing Dynamic Latent Information from DINO Features for Multi-Phase Medical Image Segmentation

Yu-Pu Hsu , Jen-Jee Chen , Yu-Chee Tseng This is my paper
classification cs.CV
keywords multi-phase CTliver lesion segmentationcontrast kineticsboundary delineationtemporal feature fusionadaptive gatingimage alignmentself-supervised features
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.

The paper tries to establish that a segmentation model which explicitly models the temporal contrast dynamics of liver lesions across multiple CT phases — how a lesion enhances in the arterial phase and washes out later — can delineate tumor boundaries more accurately than models that either use a single phase or stack phases together. The proposed framework, DynoDINO, pre-aligns slices across phases, extracts features with a medical self-supervised transformer, fuses them with a shared attention map, and then adds a gated residual of the difference between early and later phase features back into the early-phase representation. The authors report that this gated subtraction improves boundary metrics on three clinical datasets, with global HD95 dropping from about 28.7 mm to 24.4 mm on one liver-cancer cohort and from 20.2 mm to 18.5 mm on a post-treatment cohort, compared with the strongest baseline. They also report that removing the gate causes training collapse under simulated misalignment, and that the model remains accurate when the middle or late phase is missing. The reason a general reader should care is that boundary precision, not just volume overlap, is what determines safe resection margins and assessment of residual tumor.

Core claim

The central claim is that the diagnostically meaningful signal in multi-phase contrast CT is the difference between phases — the wash-in and wash-out of the contrast agent — and that a segmentation network can exploit that signal if it has a stable reference phase, a learned subtraction of the other phases from that reference, and a spatial gate that lets only well-aligned differences pass. DynoDINO aligns slices using FFT-accelerated normalized cross-correlation, encodes each phase with a shared self-supervised vision transformer, fuses phase features through a shared bidirectional attention map (Mix-attention), and then re-injects the gated difference between early and mid/late phase featu

What carries the argument

Mix-attention: a single affinity matrix computed between the reference (late-phase) stream and each target (early/mid-phase) stream, applied in both directions, so inter-phase interaction is learned with roughly half the parameters of separate cross-attention maps. Difference-based residual with adaptive gating: after fusion, early- and mid-phase features are layer-normalized, subtracted, and pushed through a sigmoid gate that produces a spatial weight map; only the gated difference is added back to the original early-phase features, which is the step that encodes wash-in/wash-out kinetics while suppressing misalignment artifacts. Supporting these is FFT-accelerated zero-mean normalized cros

Load-bearing premise

The load-bearing premise is that the early (arterial) phase is always available at inference: when it is absent, tumor segmentation collapses to near zero, so every robustness claim depends on this input being guaranteed.

What would settle it

On a motion-free, perfectly registered multi-phase liver CT dataset (e.g., phantom scans or simulated data), remove the adaptive gating mechanism; if the boundary metrics (HD95/NSD) do not degrade compared with the gated model, then the gate is only compensating for misalignment rather than preserving contrast kinetics, which would undercut the paper's central mechanistic claim.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Multi-phase segmentation systems should consider explicit phase-difference modeling rather than channel concatenation: the reported HD95 gains are consistent across three datasets with different acquisition protocols.
  • The system tolerates missing mid or late phases without retraining, so it fits clinical protocols that omit a venous or delayed phase, as long as the arterial phase is acquired.
  • Removing the adaptive gate under spatial shift leads to complete training collapse, meaning any architecture using unguided subtraction between misaligned features needs a similar safeguard.
  • The shared attention design cuts fusion parameters by about 4.1%, making the multi-phase model feasible on standard GPUs despite a large transformer backbone.
  • The authors claim the same architecture transfers to other multi-phase or multi-temporal imaging (e.g., multi-parametric MRI, DCE-MRI) without protocol-specific redesign.

Where Pith is reading between the lines

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

  • The paper's phase-dropout table shows tumor DSC collapsing to 0.59% when the early phase is missing at inference. We infer that the 'missing-phase robustness' claim is conditional: the model is robust to missing mid/late phases, not to missing the arterial reference, so clinical protocols that routinely omit the arterial phase would not benefit.
  • The shared bidirectional attention map suggests a general efficiency principle for multi-modal fusion: when two streams are strongly coupled, one affinity matrix applied in both directions may match full cross-attention; this could be tested in other fusion settings such as PET-CT or multi-modal MRI.
  • Since the gate output is a spatial weight map, a natural extension is to expose it as a per-pixel confidence map for radiologist review, turning the noise-suppression signal into an interpretable attention visualization.
  • The pre-alignment step achieves only about 65% tracking accuracy under controlled slice shifts, which implies the model's robustness under clinical conditions is substantially carried by the gating mechanism rather than by perfect alignment; a direct comparison of the gate's learned mask against known ground-truth motion would clarify the division of labor.

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 / 5 minor

Summary. The paper proposes DynoDINO, a multi-phase CECT segmentation framework that first aligns slices via FFT-accelerated ZNCC, extracts phase-wise features with a MedDINOv3 encoder, and fuses them through Mix-Attention layers followed by a difference-based residual connection with adaptive gating. The method is trained with Dice + CE loss and evaluated on LiTS, PLC-CECT, and WAW-TACE for liver/tumor segmentation. The central claim is that DynoDINO consistently improves boundary delineation and structural fidelity under standard, shifted, and missing-phase conditions, with particular emphasis on HD95 and NSD.

Significance. The intended contribution is clinically motivated: explicitly modeling inter-phase enhancement differences with a learned gate is a plausible way to capture wash-in/wash-out kinetics, and the paper provides extensive ablations, phase-dropout analysis, and a controlled misalignment experiment. No circularity was found: training uses Dice/CE on held-out data, and the gating/attention components are architectural choices rather than fitted to the target metrics. If the reported results were internally consistent, this would be a useful advance for multi-phase medical image segmentation. However, the quantitative core is currently not self-consistent, and the 'consistent improvement' claim is broader than the tables support.

major comments (3)
  1. [§4.3.2 and §4.5, Tables 4, 6, 8, 9, 10] The same configuration—DynoDINO with EP+MP+LP input—is assigned different global scores in different tables. On PLC-CECT, Table 4 reports DSC/NSD/HD95 = 78.00/47.10/24.35, while Tables 8, 9, and 10 report 76.01/47.10/24.35. On WAW-TACE, Table 6 reports 70.21/54.69/18.54, while Tables 9 and 10 report 77.33/54.15/17.79. These are not rounding artifacts: DSC differs by 1.99 and 7.12 percentage points. The text presents all rows as the same cohort and input configuration, with no note of a different split, subset, averaging scheme, or test-time protocol. Because the abstract's quantitative claims and the comparisons against nnU-Net rest on these rows, the authors must reconcile the numbers and specify the exact evaluation protocol for each table before the central claim can be assessed.
  2. [Abstract and §4.3.1, Tables 2 and 6] The claim that DynoDINO 'consistently improves boundary delineation and structural fidelity' is not supported by all reported metrics. On LiTS, DynoDINO's global NSD (49.19%) is below MedDINOv3 (52.31%), and tumor sensitivity (55.81%) is below MedDINOv3 (60.30%); the HD95 improvement over MedDINOv3 is only 0.09 mm. On WAW-TACE under single-phase input, DynoDINO's global DSC (69.72%) and tumor DSC (43.58%) are far below nnU-Net (79.04% and 62.06%), with NSD (52.10%) also below nnU-Net (54.19%). The narrative emphasizes HD95 and NSD while omitting these counterexamples. The paper should either qualify the claim to specific metrics and configurations or report all metrics in the conclusions; the current selective summary is misleading.
  3. [§3.1 exclusion criterion 2 and §4.5.3, Table 10] The 'missing-phase robustness' claim is conditional on EP being present. Section 3.1 explicitly excludes patients without EP, and Table 10 shows that removing EP at inference collapses PLC-CECT tumor DSC to 0.59% (and causes a large drop on WAW-TACE, from 77.33% to 58.22%). Thus the model is robust to missing MP/LP only, not to missing EP. This is a substantive limitation of the claimed ability to handle 'incomplete acquisitions' and should be stated in the Abstract and Conclusions. As written, the paper overstates the generality of its phase-dropout robustness.
minor comments (5)
  1. [§4.3.1] Typo: 'LPC-CECT' should be 'PLC-CECT'.
  2. [§3.2, Eq. (3)] The denominator D appears to mix elementwise products with a scalar total-energy term. Please clarify the dimensions and operations so the ZNCC normalization is unambiguous.
  3. [§3.2] The ZNCC reliability thresholds (0.75 similarity, 1.5×IQR separation) are fixed without a sensitivity analysis. Since these thresholds control which slices are padded and thereby affect input composition, a brief ablation or justification would strengthen the method.
  4. [§3.3.2 and Table 14] The text says the Mix-Attention mechanism reduces 'memory footprint' by 4.1%, but the reported numbers are parameter counts. Memory savings should be measured or the wording should be changed to 'parameter count'.
  5. [Table 13] The HD95 entry 'N/A' for the collapsed DynoDINO w/o Gating model should be explicitly defined (e.g., no foreground prediction produced), rather than left as an undefined placeholder.

Circularity Check

0 steps flagged

No circular dependency: DynoDINO's gains are empirical held-out results; reported metrics are not encoded in the model definition or training objective.

full rationale

The paper's derivation chain is non-circular: raw multi-phase CT slices are pre-aligned with FFT-based ZNCC, encoded by a pretrained MedDINOv3 backbone, fused by Mix-Attention with difference-based residual and adaptive gating, and trained with a standard Dice+CE loss. None of the reported evaluation metrics (DSC, NSD, HD95) appears in the loss, in the alignment criterion, or in the gating/attention definitions. The ZNCC reliability thresholds (0.75; 1.5×IQR) are hand-set preprocessing choices and do not encode segmentation outcomes. The phase-dropout and misalignment ablations are empirical evaluations, not fitted predictions; the EP-requirement limitation is explicitly disclosed in Table 10 and Section 4.5.3, where omitting EP collapses tumor DSC to 0.59%, which is a honest limitation rather than a circularity. The paper contains no load-bearing self-citations: the cited MedDINOv3, DINOv3, nnU-Net, and dataset sources are external. The main legitimate concern raised by the skeptical reading is an internal numerical inconsistency — the same DynoDINO multi-phase configuration is reported with different DSC values in Tables 4/6 versus Tables 8/9/10 — but that is a reproducibility/correctness issue, not a definitional reduction of a prediction to its inputs. No step in the claimed derivation reduces to its own inputs by construction, so the appropriate circularity score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The central claim depends on a set of hand-chosen preprocessing thresholds (ZNCC reliability, IQR multiplier), architecture hyperparameters (N=4 fusion layers), and domain assumptions about phase taxonomy and the interpretability of feature differences. No invented physical entities.

free parameters (3)
  • ZNCC reliability threshold = 0.75
    Hand-chosen threshold for accepting a peak ZNCC match as reliable; affects how many slices are aligned vs padded (Section 3.2, Reliability Verification).
  • IQR separation multiplier = 1.5
    Hand-chosen multiplier requiring the peak ZNCC to exceed the median by 1.5× the 75th-percentile-to-median gap; part of the false-match rejection protocol.
  • Fusion layers N = 4
    Number of iterative Mix-Attention fusion layers; chosen by the authors without sensitivity analysis (Section 3.3.2).
axioms (5)
  • domain assumption Multi-phase CECT phases can be categorized into a generalized EP/MP/LP hierarchy
    Section 3.1 states this taxonomy despite organ-specific nomenclature; if a dataset's phases don't map cleanly (e.g., only two phases in different temporal order), the method's alignment/fusion may not transfer.
  • domain assumption Layer-normalized feature differences LN(F_E)-LN(F_M) encode clinically meaningful contrast kinetics rather than alignment noise
    Eq. 11 and Section 3.3.3; the whole difference-residual design rests on this. The paper's own ablation shows without gating this subtraction can cause catastrophic collapse (Table 13), so the assumption only holds with the gate.
  • domain assumption FFT-based ZNCC is a valid similarity metric for inter-phase CT slice matching under nonlinear intensity changes
    Section 3.2; justified empirically on 900 instances with 65.25% tracking accuracy, meaning ~35% of slices are not exactly recovered.
  • standard math Convolution theorem / FFT cross-correlation
    Eq. 2-3 use F^{-1}(F(I_source) ⊙ F(I_fixed)*) for cross-correlation; standard Fourier analysis.
  • ad hoc to paper Reliability verification thresholds (0.75, 1.5×IQR) are appropriate for rejecting false matches
    No principled derivation; these constants are introduced specifically for this pipeline and could affect which slices are discarded on messy data.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DynoDINO: Harnessing Dynamic Latent Information from DINO Features for Multi-Phase Medical Image Segmentation." pith.science (2026). https://pith.science/paper/7V2H7NCC

@misc{pith2026260729568,
  author       = {Pith},
  title        = {Pith review of: DynoDINO: Harnessing Dynamic Latent Information from DINO Features for Multi-Phase Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7V2H7NCC}},
  note         = {Machine review of arXiv:2607.29568}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multi-phase Contrast-Enhanced Computed Tomography (CECT) plays a central role in the diagnosis and characterization of focal lesions by capturing temporal enhancement patterns across multiple acquisition phases. Accurate lesion segmentation from such data remains challenging because clinically relevant contrast kinetics are distributed across phases, while anatomical inconsistencies, respiratory motion, and incomplete acquisitions often lead to inter-phase misalignment and interrupted temporal information. Conventional segmentation frameworks typically process each phase independently or rely on simple fusion strategies, limiting their temporal reasoning capability. To address these challenges, we propose DynoDINO, a unified framework tailored to address the core challenges of multi-phase medical image segmentation. DynoDINO first performs slice-level alignment to establish inter-phase anatomical correspondence and then employs a Multi-phase Fusion Model to jointly enhance temporal correlations across phases. Our fusion model incorporates a Mix-attention (MA) mechanism for efficient multi-phase feature calibration and an Adaptive Gating Mechanism with difference-based residual learning to selectively preserve diagnostically relevant contrast variations while suppressing artifacts caused by residual misalignment. In addition, the adaptive gating mechanism improves training stability by preventing feature degradation caused by unguided subtraction operations. Experiments on three large-scale datasets, including LiTS, PLC-CECT, and WAW-TACE, demonstrate that DynoDINO consistently improves boundary delineation and structural fidelity under standard, shifted, and missing-phase conditions.

Figures

Figures reproduced from arXiv: 2607.29568 by Jen-Jee Chen, Yu-Chee Tseng, Yu-Pu Hsu.

Figure 1
Figure 1. Figure 1: The DynoDINO framework. Multi-phase CT slices are first aligned using the ZNCC matching method (here, [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The Fusion Module, which performs N iterations of Mix-Attention followed by a difference-based residual connection and a Sigmoid gating mechanism. The calibrated outputs are derived by applying the map in opposing directions: R′seq = RQK + Softmax  Attmap √ dk  T V (7) T ′seq = T QK + Softmax AttT map √ dk ! RV (8) 3.3.3 Difference-based Residual and Adaptive Gating After N Fusion Layers, we obtain the r… view at source ↗
Figure 3
Figure 3. Figure 3: The Mix-Attention Layer computes a shared bidirectional attention map ( [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative comparison of segmentation results on the LiTS dataset. Results are shown in red. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of segmentation results on the PLC-CECT dataset. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of segmentation results on the WAW-TACE dataset. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Analysis of training dynamics and convergence stability on the WAW-TACE dataset. Blue and red curves [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of the FFT-ZNCC alignment (Case P0001, PLC-CECT). The green and red masks represent the [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

55 extracted references · 1 canonical work pages

  1. [1]

    Magnetic resonance imaging , volume=

    Artificial intelligence in medical imaging , author=. Magnetic resonance imaging , volume=. 2020 , publisher=

    2020

  2. [2]

    Nature medicine , volume=

    A guide to deep learning in healthcare , author=. Nature medicine , volume=. 2019 , publisher=

    2019

  3. [3]

    International Conference on Medical image computing and computer-assisted intervention , pages=

    U-net: Convolutional Networks for Biomedical Image Segmentation , author=. International Conference on Medical image computing and computer-assisted intervention , pages=. 2015 , organization=

    2015

  4. [4]

    2016 fourth international conference on 3D vision (3DV) , pages=

    V-net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation , author=. 2016 fourth international conference on 3D vision (3DV) , pages=. 2016 , organization=

    2016

  5. [5]

    Journal of hepatology , volume=

    EASL Clinical Practice Guidelines: Management of hepatocellular carcinoma , author=. Journal of hepatology , volume=. 2018 , publisher=

    2018

  6. [6]

    Radiology , volume=

    Liver Imaging Reporting and Data System (LI-RADS) Version 2018: Imaging of Hepatocellular Carcinoma in At-Risk Patients , author=. Radiology , volume=. 2018 , publisher=

    2018

  7. [7]

    Cancer Imaging , volume=

    Imaging of hepatocellular carcinoma: diagnosis, staging and treatment monitoring , author=. Cancer Imaging , volume=

  8. [8]

    Best Practice & Research Clinical Gastroenterology , volume=

    Multidetector CT of hepatocellular carcinoma , author=. Best Practice & Research Clinical Gastroenterology , volume=. 2005 , publisher=

    2005

  9. [9]

    arXiv preprint arXiv:1809.10486 , year=

    nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation , author=. arXiv preprint arXiv:1809.10486 , year=

    Pith/arXiv arXiv

  10. [10]

    Array , volume=

    A review: Deep learning for medical image segmentation using multi-modality fusion , author=. Array , volume=. 2019 , publisher=

    2019

  11. [11]

    arXiv preprint arXiv:2010.11929 , year=

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale , author=. arXiv preprint arXiv:2010.11929 , year=

    Pith/arXiv arXiv 2010

  12. [12]

    Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages=

    UNETR: Transformers for 3D Medical Image Segmentation , author=. Proceedings of the IEEE/CVF winter conference on applications of computer vision , pages=

  13. [13]

    European conference on computer vision , pages=

    Swin-Unet: Unet-Like Pure Transformer for Medical Image Segmentation , author=. European conference on computer vision , pages=. 2022 , organization=

    2022

  14. [14]

    IEEE Journal of Biomedical and Health Informatics , volume=

    Is Attention all You Need in Medical Image Analysis? A Review , author=. IEEE Journal of Biomedical and Health Informatics , volume=. 2023 , publisher=

    2023

  15. [15]

    The British journal of radiology , volume=

    Enhancement patterns of hepatocellular carcinomas on multiphasic multidetector row CT: comparison with pathological differentiation , author=. The British journal of radiology , volume=. 2012 , publisher=

    2012

  16. [16]

    Abdominal Radiology , volume=

    Artificial intelligence in imaging diagnosis of liver tumors: current status and future prospects , author=. Abdominal Radiology , volume=. 2026 , publisher=

    2026

  17. [17]

    CA: a cancer journal for clinicians , volume=

    Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries , author=. CA: a cancer journal for clinicians , volume=. 2024 , publisher=

    2022

  18. [18]

    Hepatology , volume=

    Management of hepatocellular carcinoma: An update , author=. Hepatology , volume=

  19. [19]

    Hepatology , volume=

    Diagnosis, Staging, and Management of Hepatocellular Carcinoma: 2018 Practice Guidance by the American Association for the Study of Liver Diseases , author=. Hepatology , volume=. 2018 , publisher=

    2018

  20. [20]

    Hepatology , volume=

    AASLD Guidelines for the Treatment of Hepatocellular Carcinoma , author=. Hepatology , volume=. 2018 , publisher=

    2018

  21. [21]

    Machine Vision and Applications , volume=

    Deep learning in medical image registration: a survey , author=. Machine Vision and Applications , volume=. 2020 , publisher=

    2020

  22. [22]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Emerging Properties in Self-Supervised Vision Transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  23. [23]

    arXiv preprint arXiv:2509.02379 , year=

    MedDINOv3: How to adapt vision foundation models for medical image segmentation? , author=. arXiv preprint arXiv:2509.02379 , year=

    arXiv

  24. [24]

    arXiv preprint arXiv:2508.20909 , year=

    Dino U-Net: Exploiting High-Fidelity Dense Features from Foundation Models for Medical Image Segmentation , author=. arXiv preprint arXiv:2508.20909 , year=

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:1804.03999 , year=

    Attention U-Net: Learning Where to Look for the Pancreas , author=. arXiv preprint arXiv:1804.03999 , year=

    Pith/arXiv arXiv

  26. [26]

    Advances in neural information processing systems , volume=

    FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness , author=. Advances in neural information processing systems , volume=

  27. [27]

    ACM Computing Surveys , volume=

    Efficient Transformers: A Survey , author=. ACM Computing Surveys , volume=. 2022 , publisher=

    2022

  28. [28]

    Scientific reports , volume=

    Deep learning for fully automatic detection, segmentation, and Gleason grade estimation of prostate cancer in multiparametric magnetic resonance images , author=. Scientific reports , volume=. 2022 , publisher=

    2022

  29. [29]

    Magnetic Resonance Imaging Clinics , volume=

    Dynamic Contrast-Enhanced (DCE) MRI , author=. Magnetic Resonance Imaging Clinics , volume=. 2024 , publisher=

    2024

  30. [30]

    arXiv preprint arXiv:2304.07193 , year=

    Dinov2: Learning robust visual features without supervision , author=. arXiv preprint arXiv:2304.07193 , year=

    Pith/arXiv arXiv

  31. [31]

    arXiv preprint arXiv:2508.10104 , year=

    Dinov3 , author=. arXiv preprint arXiv:2508.10104 , year=

    Pith/arXiv arXiv

  32. [32]

    Medical Image Analysis , volume=

    Multi-modal learning for clinical decision support, surgery assistance and medical image analysis , author=. Medical Image Analysis , volume=. 2023 , publisher=

    2023

  33. [33]

    Advances in neural information processing systems , volume=

    SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers , author=. Advances in neural information processing systems , volume=

  34. [34]

    Image and vision computing , volume=

    Image registration methods: a survey , author=. Image and vision computing , volume=. 2003 , publisher=

    2003

  35. [35]

    IEEE transactions on medical imaging , volume=

    Voxelmorph: a learning framework for deformable medical image registration , author=. IEEE transactions on medical imaging , volume=. 2019 , publisher=

    2019

  36. [36]

    International conference on medical image computing and computer-assisted intervention , pages=

    Modality-Aware Mutual Learning for Multi-modal Medical Image Segmentation , author=. International conference on medical image computing and computer-assisted intervention , pages=. 2021 , organization=

    2021

  37. [37]

    International conference on medical image computing and computer-assisted intervention , pages=

    U-ReSNet: Ultimate Coupling of Registration and Segmentation with Deep Nets , author=. International conference on medical image computing and computer-assisted intervention , pages=. 2019 , organization=

    2019

  38. [38]

    Computer Methods and Programs in Biomedicine , volume=

    Recurrent feature fusion learning for multi-modality pet-ct tumor segmentation , author=. Computer Methods and Programs in Biomedicine , volume=. 2021 , publisher=

    2021

  39. [39]

    IEEE transactions on medical imaging , volume=

    HyperDense-Net: A Hyper-Densely Connected CNN for Multi-Modal Image Segmentation , author=. IEEE transactions on medical imaging , volume=. 2018 , publisher=

    2018

  40. [40]

    International conference on medical image computing and computer-assisted intervention , pages=

    TransFusion: Multi-view Divergent Fusion for Medical Image Segmentation with Transformers , author=. International conference on medical image computing and computer-assisted intervention , pages=. 2022 , organization=

    2022

  41. [41]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Non-local neural networks , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  42. [42]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Slowfast networks for video recognition , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  43. [43]

    IEEE Transactions on Geoscience and Remote Sensing , volume=

    Remote sensing image change detection with transformers , author=. IEEE Transactions on Geoscience and Remote Sensing , volume=. 2021 , publisher=

    2021

  44. [44]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  45. [45]

    arXiv preprint arXiv:2307.08691 , year=

    FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning , author=. arXiv preprint arXiv:2307.08691 , year=

    Pith/arXiv arXiv

  46. [46]

    Medical image analysis , volume=

    The liver tumor segmentation benchmark (lits) , author=. Medical image analysis , volume=. 2023 , publisher=

    2023

  47. [47]

    2024 , month=

    Primary Liver Cancer CECT Imaging Dataset , author=. 2024 , month=. doi:10.57760/sciencedb.12173 , publisher=

    work page doi:10.57760/sciencedb.12173 2024

  48. [48]

    Radiology: Artificial Intelligence , volume=

    WAW-TACE: a hepatocellular carcinoma multiphase CT dataset with segmentations, radiomics features, and clinical data , author=. Radiology: Artificial Intelligence , volume=. 2024 , publisher=

    2024

  49. [49]

    Vision interface , volume=

    Fast Normalized Cross-Correlation , author=. Vision interface , volume=

  50. [50]

    Optics letters , volume=

    Efficient subpixel image registration algorithms , author=. Optics letters , volume=. 2008 , publisher=

    2008

  51. [51]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  52. [52]

    BMC medical imaging , volume=

    Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool , author=. BMC medical imaging , volume=

  53. [53]

    arXiv preprint arXiv:1711.05101 , year=

    Decoupled weight decay regularization , author=. arXiv preprint arXiv:1711.05101 , year=

    Pith/arXiv arXiv

  54. [54]

    International Conference on Learning Representations (ICLR) , year=

    SGDR: Stochastic gradient descent with warm restarts , author=. International Conference on Learning Representations (ICLR) , year=

  55. [55]

    arXiv preprint arXiv:1809.04430 , year=

    Deep learning to achieve clinician-level performance in multi-organ segmentation , author=. arXiv preprint arXiv:1809.04430 , year=

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.