Pith. sign in

REVIEW 3 major objections 4 minor 19 references

SecondOpinion shows that a gate trained to predict when its own answer is wrong can invoke an anatomy-guided second stream on only 9.23% of chest X-ray cases while matching or beating prior state of the art on chest and pelvic fracture clas

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 ↗

A correctness-supervised gate invokes an anatomy-guided second stream only on hard cases, matching prior models on chest X-ray and fracture tasks with lower compute.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Clean efficiency paper with a real combination, but the headline contrast between correctness-supervised and confidence-based gating is never tested. the 3 major comments →

arxiv 2608.01808 v1 pith:5E5QXWNH submitted 2026-08-03 cs.CV

SecondOpinion: Anatomy-Aware Gated Reasoning for Efficient Medical Image Analysis

Siam Tahsin Bhuiyan , Rashedur Rahman , Sefatul Wasi , Riyadul Islam , Syoji Kobashi , Ashraful Islam , Saadia Binte Alam This is my paper
classification cs.CV
keywords conditional computationanatomy-guided learningdual-stream networkscorrectness-supervised gatingchest X-ray classificationpelvic fracture detectionearly-exit networksmedical image analysis
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 reading

Medical image classifiers typically spend the same compute on an obvious case and an ambiguous one, and dual-stream models that add anatomical guidance run both streams on every image. SecondOpinion replaces this unconditional design with a learned gate, GateKeeper, trained as a binary correctness classifier on whether the primary stream's prediction matches the label; the anatomy-guided second stream and a cross-attention fusion module run only when the gate predicts the primary answer is unreliable. On a unified five-class chest X-ray dataset the gate activates on 9.23% of cases, on visible pelvic fractures on 24.12%, and on a held-out subset of CT-confirmed but X-ray-invisible fractures on 45.71%, while accuracy and AUROC match or exceed previous state of the art. The central claim is that supervising the gate toward correctness, rather than thresholding an unsupervised confidence score, makes the model allocate anatomical reasoning where it helps most. If right, diagnostic compute scales with case difficulty rather than with the number of images.

Core claim

On its own terms, the paper's discovery is that the decision to request a second opinion can be learned as a binary correctness prediction. Phase 1 trains both streams on every sample with GateKeeper frozen; phase 2 trains the gate with binary cross-entropy against g=1[arg max(z_a)=y], and at inference the hard rule alpha<0.5 decides whether Stream B and fusion execute. On the five-class chest X-ray set SecondOpinion reaches 98.41% accuracy and 0.990 AUROC while activating the anatomy stream on 9.23% of cases, within 0.64 accuracy points of the always-on dual-stream model and above the best compared baseline (96.21%). On pelvic fracture it outperforms PelFANet in AUROC and F1 on both visible

What carries the argument

GateKeeper is a lightweight MLP that takes Stream A's pooled features, its softmax probabilities, and the Shannon entropy of those probabilities, and outputs alpha, a predicted probability that Stream A's argmax is correct. It is trained with binary cross-entropy against the true correctness label, so the natural decision boundary is 0.5; the paper uses no tuned threshold and no efficiency-penalty term. Stream B is a second, independently weighted EfficientNet-B0 whose input is a predicted segmentation map rather than the raw image, so it contributes anatomy evidence with no shared intermediate features. When alpha<0.5, a single-head cross-attention fusion module uses Stream A as query and S

Load-bearing premise

For the chest X-ray results, the tuberculosis Stream B inputs are inferred from a segmentation model trained on the other four classes, and the paper does not state whether that model was trained per-fold or on all non-TB data; if the latter, the TB test cases' anatomical inputs came from a model that saw test-fold images of other classes, contaminating the five-fold cross-validation and inflating Stream B's contribution.

What would settle it

Two checks would settle the central claim: (1) retrain the tuberculosis segmentation model strictly inside each training fold and rerun the five-fold chest X-ray evaluation, and (2) replace GateKeeper on the same Stream A with the best threshold on max-softmax probability or entropy; if the TB results degrade substantially, or if a simple confidence threshold reproduces the same activation/accuracy curve, the correctness-supervised gate is not what the paper claims it to be.

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

If this is right

  • Chest X-ray classification runs at 0.442 GFLOPs and 4.78M parameters with 9.23% activation, roughly half the always-on cost, while staying within 0.64 accuracy points of the unconditional dual-stream model.
  • The same unmodified gate spends 24.12% of its activations on visible fractures and 45.71% on invisible fractures, showing that compute automatically rises when the input distribution shifts away from what Stream A can resolve.
  • No confidence threshold needs tuning: the binary cross-entropy decision boundary at alpha=0.5 is the only operating point, and the gate's objective targets correctness rather than efficiency, with efficiency following as a consequence.
  • Because Stream B and fusion run only on activated cases, adding anatomy guidance does not double inference cost for the majority of inputs; the measured gap to the always-on model (0.6-2.2 accuracy points) is the explicit cost of that efficiency.

Where Pith is reading between the lines

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

  • GateKeeper's activation rate looks like a cheap, model-internal measure of dataset difficulty or distribution shift: the same trained system moved from 9.23% activation on chest X-ray to 45.71% on CT-confirmed invisible fractures, so a deployment site could monitor activation rate as a signal that the population differs from training.
  • The gate output could double as a referral signal: cases with alpha near or below 0.5 are precisely the ones Stream A is unsure about, so a clinician could use those cases for CT follow-up or specialist review; the invisible-fracture numbers (45.71% flagged) suggest this is a plausible triage tool, though the paper does not test it.
  • The two-phase, correctness-supervised recipe is not tied to X-rays or to segmentation: any task with an anatomical or costly second opinion (an atlas-based stream, a retrieval database, or a more expensive model) could use the same gate, since the supervision signal is just 'was the primary answer right'.
Share X Bluesky LinkedIn Reddit HN

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 / 4 minor

Summary. The paper proposes SecondOpinion, a conditional dual-stream framework for medical image classification. A fast primary stream (EfficientNet-B0 on raw images) processes every case; GateKeeper, a small MLP trained as a binary correctness classifier, decides whether to activate a second anatomy-guided stream (EfficientNet-B0 on predicted segmentations) and a cross-attention fusion module. Training uses a two-phase procedure with a soft-gated objective (Eq. 5) and hard gating at inference. The method is evaluated on a unified five-class chest X-ray dataset (21,865 images) and a private pelvic fracture dataset including a held-out invisible-fracture subset. The authors report that SecondOpinion matches or exceeds prior state-of-the-art accuracy/AUROC while activating the second stream on 9.23% of chest X-ray cases, 24.12% of visible fractures, and 45.71% of invisible fractures, and they attribute this behavior to correctness-supervised gating rather than unsupervised confidence.

Significance. If fully validated, the paper makes a useful contribution: it demonstrates that conditional computation can be combined with anatomy-guided feature fusion, and it provides a clean internal comparison (Stream A only, Stream B only, Always-On, SecondOpinion) on two tasks with different difficulty gradients. The inclusion of a harder, CT-confirmed invisible-fracture subset is a valuable testbed, and the activation-weighted FLOPs reporting makes the efficiency gain concrete. The central mechanistic claim, however, is not actually tested against the alternative it is explicitly distinguished from: no confidence-threshold or entropy-threshold gate baseline appears anywhere. The TB segmentation protocol also raises a potential data-contamination concern that the manuscript does not resolve. The framework and results are promising, but the stated mechanism and one of the two headline evaluations need additional support before the claims can be accepted.

major comments (3)
  1. [Abstract; §4.2, Tables 1–4] The paper's central claim is that supervising GateKeeper as a binary correctness classifier, 'rather than relying on unsupervised confidence,' is what enables efficient allocation of the anatomy-guided stream. This claim is never directly tested: there is no baseline in which Stream B is activated by a threshold on Stream A's softmax confidence or entropy. Since GateKeeper's input already includes softmax probabilities and entropy (Eq. 2), a simple entropy/confidence gate is a natural, cheaper baseline. Add such a baseline (with threshold sweep) and report the same accuracy/AUROC/activation-rate metrics. Without this control, the reported activation rates (9.23%, 24.12%, 45.71%) and the statement that activation 'tracks difficulty' are equally consistent with an unsupervised confidence gate, leaving the mechanism claim unsupported.
  2. [§2.2, §3.1] The tuberculosis Stream B input is stated to be 'inferred from the model trained on the other four classes,' but the manuscript never states whether that segmentation model was trained per-fold or on the full non-TB set. Because chest X-ray results use stratified 5-fold cross-validation, training the segmentation model on all non-TB images would mean it saw images from the held-out folds (of the other classes), contaminating the evaluations of TB test cases and potentially inflating Stream B's contribution. Please specify the exact protocol; if the segmentation model was not trained per-fold, re-run the chest X-ray evaluation with a leakage-free per-fold segmentation model and report whether the conclusions change.
  3. [§3.1, Table 3] The INVIS subset is described as a 'held-out' generalization set, but results are reported as 'Mean over 5-fold CV' with 95% CIs. This is ambiguous: if the 35 invisible-fracture cases were excluded from training for all folds, the cross-validation description is misleading; if they were used in the test split of each fold, then they were not fully held out. Clarify how the held-out status and the 5-fold averaging are reconciled, since this directly affects the strength of the generalization claim and the meaning of the reported confidence intervals.
minor comments (4)
  1. [Introduction, §4.3] There is a typo in the Introduction: 'discoveredunsupervised[2]' should be 'discovered unsupervised [2]'. Also, the GradCAM analysis is described as a 'sanity check' in the Limitations, which is appropriate, but the two-case discussion in §4.3 should be more clearly labeled as illustrative rather than evidential.
  2. [§3.3, Tables 2–3] The metric called F1 for the pelvic fracture experiments is actually the harmonic mean of Recall and Specificity, not the standard harmonic mean of Precision and Recall. Since this is a non-standard definition, please rename it (e.g., 'F1-like score') or also report the conventional F1, even if prior work used the current convention.
  3. [Table 4] The 'activation-weighted average Params' column is conceptually odd: model parameters are typically loaded whether or not a path is executed. The FLOPs values are meaningful, but the parameter values should either be stated as static per-path costs or replaced with an estimate of deployed memory under conditional loading.
  4. [§2.5] Equation (5) includes the term BCE(α, ĝ) but the paper does not specify whether the auxiliary CE terms are computed on the soft-gated prediction only, or on each stream separately. The text says 'unweighted four-term objective' but the first three terms are all CE losses; please list the exact arguments of each term to avoid ambiguity.

Circularity Check

1 steps flagged

Activation-rate-difficulty trend is a by-construction consequence of the correctness-supervised gate; the central contrast with unsupervised confidence is asserted but never tested.

specific steps
  1. fitted input called prediction [Section 2.3, GateKeeper after Eq. (5); Section 4.2 and Contribution 4]
    "This objective targets correctness prediction, not efficiency directly; efficiency follows as a consequence, since a well-calibrated gate activates Stream B only where Stream A is unreliable."

    GateKeeper is trained with binary cross-entropy against the correctness label g=1[argmax(z_a)=y] (Eq. 5). Its activation rate is therefore a direct readout of its predicted error probability for Stream A, not an independent discovery. The paper's reported monotonic activation increase across tasks (9.23% to 24.12% to 45.71%) is a statistical consequence of fitting a correctness classifier: harder subsets have lower Stream A accuracy, so a calibrated gate must activate more often there. Presenting this by-construction trend as evidence that correctness-supervised gating allocates anatomical reasoning where needed is circular, because the same trend would appear for any well-calibrated confidence or entropy threshold. No baseline using unsupervised confidence gating is run, so the claimed ad

full rationale

This is an empirical architecture paper with no analytic derivation chain, so there is no equation-level circularity. The main empirical result—matching or exceeding prior SOTA while activating the second stream only fractionally—is self-contained and not forced by construction. The most notable self-referential element is that nearly all prior-SOTA baselines (refs 1, 2, 3, and 9) are the authors' own prior work; however, these are published, externally checkable numbers, so this is self-citation rather than circularity under the stated rules. The one mild circular step is the framing of GateKeeper's activation-rate-versus-difficulty trend as an emergent finding: because GateKeeper is explicitly trained to predict Stream A's correctness (Section 2.3, Eq. 5), its activation rate is a direct function of the training target, and the difficulty-correlated activation pattern is a by-construction property rather than independent evidence. Separately, the paper's central mechanism claim—that supervising toward correctness is better than relying on unsupervised confidence—is never tested against any confidence/entropy-threshold baseline; this is a missing control that weakens the claim, but it is not itself circularity. The Limitations section acknowledges single-backbone and small-dataset issues but does not list this omitted comparison. Overall, the central performance/efficiency trade-off is independently validated, so the circularity score is modest: 3 rather than higher.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The framework rests on four domain assumptions (transfer learning, segmentation quality, gate sufficiency, statistical approximation) and one ad hoc training design (unweighted four-term loss). No new physical entities are introduced. The free parameters are standard hyperparameters plus the 0.5 gate threshold; none is fitted to the reported outcome metrics in a way that breaks the evaluation, but the TB mask inference assumption is the least protected.

free parameters (5)
  • Phase 1/2 learning rate = 1e-4
    Chosen by hand for both phases (Section 3.2); standard AdamW rate, not tuned per task.
  • Batch size = 16
    Fixed across tasks (Section 3.2).
  • Epoch schedule = CXR 25/35, pelvic 15/25
    Chosen per dataset scale (Section 3.2).
  • Fracture/normal augmentation ratios = 2x fracture, 6x normal
    Chosen to counter class imbalance in the 228-image pelvic dataset (Section 3.1).
  • Gate activation threshold = 0.5
    Natural BCE decision boundary, not tuned (Section 2.3); still a design choice that sets the efficiency-accuracy tradeoff.
axioms (5)
  • domain assumption ImageNet-pretrained EfficientNet-B0 weights transfer to chest X-ray and pelvic radiograph features.
    Invoked in Section 2.1 for Stream A and implicitly for Stream B; standard transfer learning but not proven for these tasks.
  • domain assumption A U-Net (MiT-B0) trained on four chest X-ray classes produces segmentation masks useful for the fifth class (tuberculosis).
    Section 2.2 uses inferred TB masks as Stream B input; the quality of these masks on unseen TB images is never evaluated.
  • domain assumption Stream A's pooled features, softmax probabilities, and entropy are sufficient for GateKeeper to predict Stream A's correctness.
    Section 2.3 defines the gate input; no calibration analysis or feature importance is provided.
  • domain assumption 5-fold cross-validation with z-approximation 95% CIs is statistically valid at 228 pelvic images and 35 invisible-fracture cases.
    Section 3.3 and Limitations; the z-approximation is questionable at n=35.
  • ad hoc to paper The soft-gated training objective (Equation 5) with unweighted terms trains the gate without a separate efficiency constraint.
    Section 2.5; the absence of an efficiency penalty is a design choice, not a proven property.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SecondOpinion: Anatomy-Aware Gated Reasoning for Efficient Medical Image Analysis." pith.science (2026). https://pith.science/paper/5E5QXWNH

@misc{pith2026260801808,
  author       = {Pith},
  title        = {Pith review of: SecondOpinion: Anatomy-Aware Gated Reasoning for Efficient Medical Image Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5E5QXWNH}},
  note         = {Machine review of arXiv:2608.01808}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Deep learning models for medical image analysis typically apply a fixed amount of computation to every input, regardless of case difficulty. Anatomy-guided dual-stream architectures have been shown to improve diagnostic performance, but they evaluate both streams unconditionally, even on cases a single stream could already resolve confidently. We propose SecondOpinion, a framework in which a fast primary stream processes every case, while a second, anatomy-guided stream is invoked only when GateKeeper, a gating mechanism trained explicitly as a binary correctness classifier, judges that the primary stream's prediction needs additional scrutiny, much as a clinician might seek a second opinion on a difficult case. When activated, the two streams are combined through a lightweight cross-attention fusion module. We evaluate SecondOpinion on a unified five-class chest X-ray dataset and a pelvic fracture dataset, the latter including a held-out, harder subset of fractures that are invisible on X-ray but confirmed via CT. SecondOpinion matches or exceeds prior state-of-the-art performance on both tasks, while activating its anatomy-guided stream on only 9.23% of chest X-ray cases, rising to 24.12% on visible fractures and 45.71% on invisible fractures, an activation rate that tracks task difficulty directly. These results suggest that supervising a gating signal toward correctness, rather than relying on unsupervised confidence, allows a model to allocate anatomical reasoning where it is actually needed.

Figures

Figures reproduced from arXiv: 2608.01808 by Ashraful Islam, Rashedur Rahman, Riyadul Islam, Saadia Binte Alam, Sefatul Wasi, Siam Tahsin Bhuiyan, Syoji Kobashi.

Figure 1
Figure 1. Figure 1: SecondOpinion Full Pipeline. SecondOpinion processes every input through a fast primary stream, and invokes a second, anatomy-guided stream only when a learned gating mecha￾nism, GateKeeper, flags the primary stream’s prediction as uncertain [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: GradCAM examples. (a) Chest X-ray (Stream B inactive): Stream A attends correctly to both lungs. (b) Invisible fracture (Stream B active): Stream A attends to irrelevant background corners. (c) Same case post-fusion: attention refocuses onto the bone region. rectly to both lung fields. For an invisible fracture case (Stream B active), Stream A instead attends to irrelevant background corners; after fusion,… 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

19 extracted references · 2 linked inside Pith

  1. [1]

    Soft Computing29(23), 6333–6345 (2025)

    Bhuiyan, S.T., Rahman, R., Islam, R., Haque, M., Islam, M., Kobashi, S., Alam, S.B.: Lung disease diagnosis from cxr using convolutional neural network (cnn): A comparative study. Soft Computing29(23), 6333–6345 (2025)

    2025

  2. [2]

    In: Workshop on Empowering Medical Image Computing and Research through Early-Career Expertise

    Bhuiyan, S.T., Rahman, R., Wasi, S., Yagi, N., Kobashi, S., Islam, A., Alam, S.B.: Invisible yet detected: Pelfanet with attention-guided anatomical fusion for pelvic fracture diagnosis. In: Workshop on Empowering Medical Image Computing and Research through Early-Career Expertise. pp. 103–113. Springer (2025)

    2025

  3. [3]

    In: 2026 IEEE 56th International Sym- posium on Multiple-Valued Logic (ISMVL)

    Bhuiyan, S.T., Zaman, Z., Rahman, R., Choudhury, M.S., Wasi, S., Islam, A., Alam,S.B.,Kobashi,S.:Chexnet-cbam:Enhancinglarge-scalemedicalpre-training with attention for lung disease diagnosis. In: 2026 IEEE 56th International Sym- posium on Multiple-Valued Logic (ISMVL). pp. 105–110. IEEE (2026)

    2026

  4. [4]

    Journal of Medical Systems 49(1), 162 (2025)

    Bruno, P., Macrì, M., Dodaro, C.: A dual-stage deep learning framework for breast ultrasound image segmentation and classification. Journal of Medical Systems 49(1), 162 (2025)

    2025

  5. [5]

    Chowdhury, M.E., Rahman, T., Khandakar, A., Mazhar, R., Kadir, M.A., Mahbub, Z.B., Islam, K.R., Khan, M.S., Iqbal, A., Al Emadi, N., et al.: Can ai help in screening viral and covid-19 pneumonia? Ieee Access8, 132665–132676 (2020)

    2020

  6. [6]

    In: Medical Imaging 2022: Image Processing

    Jin, C., Udupa, J., Zhao, L., Tong, Y., Odhner, D., Pednekar, G., Nag, S., Lewis, S., Poole, N., Mannikeri, S., et al.: Anatomy-guided deep learning for object local- ization in medical images. In: Medical Imaging 2022: Image Processing. vol. 12032, pp. 591–597. SPIE (2022)

    2022

  7. [7]

    arXiv preprint arXiv:2509.17885 (2025)

    Mokssit, S., Karrakchou, O., Mousist, A., Ghogho, M.: Confidence-gated training for efficient early-exit neural networks. arXiv preprint arXiv:2509.17885 (2025)

    arXiv 2025

  8. [8]

    Pattern Recog- nition105, 107346 (2020)

    Passalis, N., Raitoharju, J., Tefas, A., Gabbouj, M.: Efficient adaptive inference for deep convolutional neural networks using hierarchical early exits. Pattern Recog- nition105, 107346 (2020)

    2020

  9. [9]

    Scientific Reports14(1), 8004 (2024) 10 S

    Rahman, R., Yagi, N., Hayashi, K., Maruo, A., Muratsu, H., Kobashi, S.: Enhanc- ing fracture diagnosis in pelvic x-rays by deep convolutional neural network with synthesized images from 3d-ct. Scientific Reports14(1), 8004 (2024) 10 S. T. Bhuiyan et al

    2024

  10. [10]

    Ieee Access8, 191586–191601 (2020)

    Rahman, T., Khandakar, A., Kadir, M.A., Islam, K.R., Islam, K.F., Mazhar, R., Hamid, T., Islam, M.T., Kashem, S., Mahbub, Z.B., et al.: Reliable tuberculosis detection using chest x-ray with deep learning, segmentation and visualization. Ieee Access8, 191586–191601 (2020)

    2020

  11. [11]

    Computers in biology and medicine132, 104319 (2021)

    Rahman, T., Khandakar, A., Qiblawey, Y., Tahir, A., Kiranyaz, S., Kashem, S.B.A., Islam, M.T., Al Maadeed, S., Zughaier, S.M., Khan, M.S., et al.: Exploring the effect of image enhancement techniques on covid-19 detection using chest x-ray images. Computers in biology and medicine132, 104319 (2021)

    2021

  12. [12]

    arXiv preprint arXiv:1711.05225 (2017)

    Rajpurkar, P., Irvin, J., Zhu, K., Yang, B., Mehta, H., Duan, T., Ding, D., Bagul, A., Langlotz, C., Shpanskaya, K., et al.: Chexnet: Radiologist-level pneumonia de- tection on chest x-rays with deep learning. arXiv preprint arXiv:1711.05225 (2017)

    Pith/arXiv arXiv 2017

  13. [13]

    In: Proceedings of the IEEE international conference on computer vision

    Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad- cam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. pp. 618–626 (2017)

    2017

  14. [14]

    The Bell system tech- nical journal27(3), 379–423 (1948)

    Shannon, C.E.: A mathematical theory of communication. The Bell system tech- nical journal27(3), 379–423 (1948)

    1948

  15. [15]

    In: International conference on machine learning

    Tan, M., Le, Q.: Efficientnet: Rethinking model scaling for convolutional neural networks. In: International conference on machine learning. pp. 6105–6114. PMLR (2019)

    2019

  16. [16]

    arXiv preprint arXiv:2604.17360 (2026)

    Tang, Z., Zhao, S.: T-dumpra: Teacher-guided dual-path multi-prototype retrieval augmented framework for fine-grained medical image classification. arXiv preprint arXiv:2604.17360 (2026)

    Pith/arXiv arXiv 2026

  17. [17]

    Advances in neural information pro- cessing systems30(2017)

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017)

    2017

  18. [18]

    Ad- vances in Neural Information Processing Systems34, 2516–2528 (2021)

    Wołczyk, M., Wójcik, B., Bałazy, K., Podolak, I.T., Tabor, J., Śmieja, M., Trzcin- ski, T.: Zero time waste: Recycling predictions in early exit neural networks. Ad- vances in Neural Information Processing Systems34, 2516–2528 (2021)

    2021

  19. [19]

    Journal of Imaging11(4), 101 (2025)

    Zafar, J., Koc, V., Zafar, H.: Dual-stream contrastive latent learning generative adversarial network for brain image synthesis and tumor classification. Journal of Imaging11(4), 101 (2025)

    2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.