Pith. sign in

REVIEW 4 major objections 4 minor 84 references

Reasoning in machine vision by learning fast and slow thinking

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper claims that a machine can reason in vision by spending extra inference-time compute to iteratively refine its own segmentations, and that with only 8–16 labelled samples this 'System II' process sets state-of-the-art cancer…

desk verdict Strong empirical evidence that inference-time refinement helps few-shot segmentation, but the reward mechanism is unvalidated and the SOTA/human-expert claims outrun the protocol. read the letter →

arxiv 2506.22075 v2 pith:T46XYPNH submitted 2025-06-27 cs.CV

classification cs.CV
keywords machinereasoninginference-timecomputedual-processcognitionmedicalimagesegmentationcancerlocalisationself-playreinforcementlearningfew-shotadaptationdistributiondiscriminator
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

The paper argues that a machine vision system can improve its own answers by thinking longer at inference time, without more labelled data, and that this 'slow thinking' can beat both conventional supervised learning and human experts. The proposed design pairs a fast predictor (System I) with a slow refinement module (System II) that iteratively proposes competing pixel-level edits to a segmentation and keeps the edits that a learned discriminator scores as more human-like. On cancer localisation in five organs, the authors report that System II refinement, starting from only 8–16 labelled samples, outperforms deep-learning baselines, foundation models fine-tuned on hundreds of samples, and, for prostate cancer detection, radiologists. The paper's positive claim is that inference-time compute is a practical substitute for labelled data in non-verbal vision tasks.

What carries the argument

The carrying mechanism is the 'auto-competing' refinement loop, a self-play reinforcement-learning process in which two networks compete to refine a segmentation and a learned distribution-discriminator supplies the reward. The distribution-discriminator is a network trained adversarially alongside the task-predictor; it scores an image-label pair by how likely the label is to come from human rather than machine output, and after few-shot adaptation it estimates correctness for the new task. The competitors propose small binary flips, the higher-scoring proposal is kept, and the reward from Equation (4) trains the competitors through Equation (5) until convergence, at which point the sample counts as solved. This mechanism is what converts added inference-time compute into improved segmentation.

What would settle it

Take the five cancer held-out sets, and after each System II iteration record both the discriminator score and the Dice against ground truth for the same sample. If a meaningful fraction of accepted refinements raise the discriminator score while lowering Dice, the reward signal is not aligned with correctness and the central claim fails; the same design with the discriminator replaced by a random score should also be checked to bound the effect of refinement alone.

Watch

Extended reading notes

Core claim

The central claim is that extended inference-time computation, implemented as an auto-competing refinement process, is itself a mechanism for improving performance in vision, and that it can surpass what the authors obtain from large supervised datasets, foundation models, and human experts. A System I module, adversarially trained across abdomino-pelvic CT/MR tasks and adapted with 8–16 labels, provides both an initial segmentation and a distribution-discriminator that scores how plausible a segmentation is. A System II module then runs two competing refinement networks: each proposes a small pixel flip, the discriminator scores both proposals, the higher-scoring refinement is kept, and the competitors are updated by self-play reinforcement learning to keep winning. This loop continues until the discriminator score stops improving, after which the solved sample is fed back to retrain System I and reduce future thinking time. The authors report that this yields state-of-the-art Dice for prostate, liver, pancreas, colon, and kidney cancer segmentation, and sensitivity and specificity of 0.911 and 0.655 against radiologists' 0.88 and 0.45 on histopathology-confirmed prostate cancer detection.

Load-bearing premise

The load-bearing premise is that the distribution-discriminator, adapted with only 8–16 labelled samples, gives a reward that tracks true segmentation quality; the paper never reports the correlation between discriminator scores and held-out Dice, so if the discriminator can be exploited, System II could appear to improve while actually degrading accuracy.

Editorial extensions

If this is right

  • If the central claim holds, adding compute at test time is a general lever for accuracy in dense vision tasks, independent of more labels.
  • The same System II loop applies to any task whose solutions live in a structured space where small edits can be scored, including detection and classification; the paper demonstrates a classification variant on noisy digits.
  • Methods that currently compete by scaling training data or model parameters could instead be compared on an inference-time compute budget, since System II's performance curve keeps rising with thinking time.
  • The System I retraining step means a deployment can become faster over time on a fixed task, as solved samples gradually reduce the amount of System II refinement needed.

Reading between the lines

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

  • A decisive test the paper leaves implicit: measure per-sample correlation between discriminator score and held-out Dice; if the discriminator's score can be inflated by edits that genuinely reduce segmentation quality, the reward signal is gameable and the reported improvements would not transfer to new tasks.
  • The same machinery could in principle be attached to any frozen predictor, foundation model included, turning a single forward pass into an iterative refinement; the paper mentions this as future potential in the conclusion.
  • If the reported pattern generalizes, inference-time compute budgets become a new axis of comparison for medical imaging AI, and the bottleneck shifts from annotation cost to GPU hours, an economic trade-off the paper's carbon-footprint table begins to quantify.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a dual-process framework for vision tasks, inspired by System I/System II cognition, in which a meta-trained task predictor and distribution discriminator are adapted with 4–16 labelled samples, and a System II module then refines each test-time prediction via a self-play reinforcement learning process whose reward is the discriminator score. The authors claim that this inference-time refinement converts additional compute into improved segmentation accuracy even with very few labels, and they report results on noisy digit segmentation, ImageNet-S object segmentation, five cancer localisation tasks (prostate, liver, pancreas, colon, kidney), and a patient-wise prostate cancer detection comparison against published radiologist performance.

Significance. If the central claim holds, the paper would be significant: it suggests a general mechanism for improving vision segmentation with inference-time compute in data-scarce settings, which is relevant to medical imaging and other domains where labels are expensive. The paper has genuine strengths: all headline Dice results are computed against external held-out ground truth, so the main numbers are not circular; the evaluation spans multiple real clinical tasks; the authors provide ablations (Table 4), an uncertainty analysis, a carbon-cost table, and state that code is open-sourced. However, the load-bearing mechanism—that maximising the adapted discriminator score moves segmentations toward true ground truth—is never directly validated. Because the reward is learned from the same few labels and the refinement hyperparameters are selected on the target tasks, the current evidence cannot distinguish a genuinely useful reasoning process from an incidental denoising or reward-exploitation effect, despite the plausibility of the final held-out results.

major comments (4)
  1. [Results, Table 3 and Fig. 5c] The reward used to train the competing refinement networks is the adapted distribution-discriminator score d(x,y;θ*), but the paper never measures whether this score is positively correlated with ground-truth agreement on held-out samples. The only reported evidence is the final Dice after convergence (Table 3) and the ablation in Table 4, neither of which isolates reward alignment from other effects such as incidental smoothing or per-sample fine-tuning. This is load-bearing because the central claim is that iterative refinement guided by the discriminator improves true segmentation accuracy. I request a direct correlation analysis: on held-out samples, compute Dice against ground truth and the discriminator score for intermediate refinement trajectories; a positive, monotonic relationship is necessary to support the mechanism. A control with a non-informative or random reward would also help.
  2. [Methodological details, 'The definition of refinement' and hyperparameters] The claim that 'System II thinking achieves its peak performance using only 8 labelled samples' is based on selecting the best adaptation count per task from {8,12,16}: prostate peaks at 12, pancreas at 16, liver and colon at 8, kidney at 8. Reporting per-task peaks when comparing against other methods (including the >100-label baselines) selects the most favourable configuration without accounting for the three-way choice, which inflates the apparent advantage and the strength of the 8-16 label claim. Please either prespecify the adaptation count for each task, or report all three configurations in every comparison and discuss the multiple-testing issue.
  3. [Results, 'System II thinking outperforms radiologists'] The refinement parameters p=0.01 and iend=p×P×100 are stated to be fixed by a grid search across the five evaluation applications, meaning that at least two free parameters are selected using the target tasks themselves. This is not reported in the main text or the supplementary tables; the grid range, the selection criterion, and the sensitivity of the held-out results to p and iend should be given. Without this, the data-efficiency claim is weakened because part of the improvement may come from task-specific parameter tuning on the test applications.
  4. [Results, 'System II thinking sets state-of-the-art performance'] The comparison against human experts uses published sensitivity (0.88) and specificity (0.45) from the PROMIS study rather than an independent evaluation of radiologists on the same held-out set with the same case mix. The patient-wise classification of the System II model is derived from zonal predictions, and the confidence intervals are reported for System II but not re-estimated for the radiologist numbers. Please clarify whether the radiologist figures are directly comparable on the same patients, and if not, temper the 'outperforms radiologists' claim accordingly.
minor comments (4)
  1. [Abstract and Results, 'System II thinking sets state-of-the-art performance'] The abstract states that System II achieves superior performance 'across all five organs', but the kidney comparison with ASeg is statistically equivalent (p=0.051). The wording should be adjusted to 'superior or equivalent' to match the reported data.
  2. [Supplementary, Algorithm 2] Algorithm 2 contains typos: the lines computing the distribution score for competitor b re-use the argument y_{i=0+1,a} instead of y_{i=0+1,b}, and the assignment in the 'If' branch sets y_{i=0+1}=y_{i=0+1,b} rather than y_{i=0+1,a}. Please correct these inconsistencies.
  3. [Supplementary, Eq. (5)] Eq. (5) uses \phi both as the generic weights and as the optimal solution \phi^*, and the expectations are over yi,a and yi,b whose conditional dependence on the current label is not made explicit. Clarifying the notation would improve reproducibility.
  4. [Methodological details, hyperparameters] For a 256×256 image, iend=p×P×100 equals 65,536 refinement iterations per sample; this large number should be stated in the main text so that the per-sample computational cost is transparent, and the GPU-hour table should be reformatted to make the columns legible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central compute-time improvement claims are validated against external held-out Dice and histopathology-verified benchmarks, not against the paper's own discriminator reward.

full rationale

The paper's headline results are measured with external ground truth: Dice on held-out splits of the five cancer tasks, sensitivity/specificity against histopathology, and comparisons to previously published DL and foundation-model baselines. These metrics are not defined by the System II reward, so the central claim does not reduce to the method's own inputs. The System II reward is a learned discriminator score (Eqs. 4-5), and the paper never directly validates that maximizing this score correlates with Dice; that is a genuine correctness risk (reward hacking / proxy misalignment), but it is not circular because the reported improvements are externally evaluated. The stopping/convergence criteria (Eqs. 6-7) are defined in terms of the discriminator and reward sums, but the paper does not use those criteria as the evidence for accuracy gains. Self-citations to [47] and [58] provide building blocks (self-play refinement and pre-training scheme), but they are not invoked as proof of the new claim; the new claim is supported by held-out benchmark comparisons. The pseudo-labelling loop from 'solved' samples to System I is also empirically assessed with test-set performance rather than asserted by construction. Therefore no specific equation-level or definition-level circularity can be exhibited.

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

The central claims rest on transfer from meta-learned organ segmentation, on the validity of the learned discriminator as a reward, and on fixed refinement hyperparameters that were tuned on the evaluation tasks themselves.

free parameters (3)
  • p (proportion of pixels to flip per refinement iteration) = 0.01
    Set by grid search across the five cancer localisation tasks that are also used for evaluation; this is a fitted hyperparameter affecting all segmentation results.
  • iend (terminal iteration count) = p * P * 100 (65536 for 256x256 images)
    Derived from p and used to stop the refinement loop; not independently justified and co-tuned with p.
  • Adaptation sample count per task = 8, 12, or 16, with the peak chosen per task
    The paper reports 'System II performances at their peaks', selecting the best of three sample counts separately for each organ, which is a post-hoc selection.
assumptions (4)
  • domain assumption Reptile meta-learning across 13 anatomical structures transfers to unseen cancer segmentation tasks.
    The System I module is pretrained on organ segmentation with Reptile, and the paper assumes this provides adaptable representations for tumour localisation.
  • ad hoc to paper The distribution-discriminator, adapted with a few labels, provides a reward signal whose maximization improves true segmentation accuracy.
    This is the core reward mechanism in Eq. 4-5; the paper never directly verifies the correlation between discriminator scores and held-out Dice.
  • ad hoc to paper Flipping a fixed proportion of pixels (p=0.01) per iteration is a sufficient action space for refining segmentations.
    The refinement space is defined as XOR with a rounded mask of p pixels; this is a strong restriction on the kind of corrections the system can make.
  • domain assumption Histopathology template-mapping biopsy labels are a reliable ground truth for prostate cancer localisation.
    The human-expert comparison uses these labels to define cancer presence, and the paper assumes they are accurate enough for the reported sensitivity and specificity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reasoning in machine vision by learning fast and slow thinking." pith.science (2026). https://pith.science/paper/T46XYPNH

@misc{pith2026250622075,
  author       = {Pith},
  title        = {Pith review of: Reasoning in machine vision by learning fast and slow thinking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T46XYPNH}},
  note         = {Machine review of arXiv:2506.22075}
}
read the original abstract

Reasoning is a hallmark of human intelligence, enabling adaptive decision-making in complex unfamiliar scenarios. In contrast, machine intelligence remains bound to training data, unable to dynamically refine solutions at inference. While recent advances have explored machine reasoning - trading inference-time compute for improved performance - they focus on verbal domains such as mathematical problem-solving where explicit rules govern step-by-step solution generation. Many tasks lack sufficient labelled data and require alternative performance improvement mechanisms, such as inference-time compute. Here we present a paradigm for machine reasoning in vision, enabling performance improvements with increasing thinking time (inference-time compute), even with limited labelled data. Our approach is inspired by dual-process theories of human cognition, integrating a fast-thinking System I module for generating and verifying solutions in familiar tasks, with a slow-thinking System II module that iteratively refines predictions using self-play reinforcement learning, even when task-specific data is limited. This paradigm involves proposing, competing over, and refining solutions until convergence. We demonstrate that extended inference-time compute yields superior performance compared to large-scale supervised learning, foundation models, and human experts in vision tasks. These include computer-vision benchmarks and cancer localisation across five organs, highlighting the potential of inference-time compute for data-scarce problems.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

84 extracted references · 52 canonical work pages

  1. [1]

    American economic review 93(5), 1449–1475 (2003)

    Kahneman, D.: Maps of bounded rationality: Psychology for behavioral eco- nomics. American economic review 93(5), 1449–1475 (2003)

  2. [2]

    British Journal of Psychology 75(4), 451–468 (1984)

    Evans, J.S.B.: Heuristic and analytic processes in reasoning. British Journal of Psychology 75(4), 451–468 (1984)

  3. [3]

    Scientific American 15 (2012)

    Kahneman, D.: Of 2 minds: How fast and slow thinking shape perception and choice [excerpt]. Scientific American 15 (2012)

  4. [4]

    Advances in neural information processing systems 30 (2017)

    Anthony, T., Tian, Z., Barber, D.: Thinking fast and slow with deep learning and tree search. Advances in neural information processing systems 30 (2017)

  5. [5]

    Trends in cognitive sciences 7(10), 454–459 (2003)

    Evans, J.S.B.: In two minds: dual-process accounts of reasoning. Trends in cognitive sciences 7(10), 454–459 (2003)

  6. [6]

    Medical image analysis 42, 60–88 (2017)

    Litjens, G., Kooi, T., Bejnordi, B.E., Setio, A.A.A., Ciompi, F., Ghafoorian, M., Van Der Laak, J.A., Van Ginneken, B., S´ anchez, C.I.: A survey on deep learning in medical image analysis. Medical image analysis 42, 60–88 (2017)

  7. [7]

    Therapeutic advances in urology 14, 17562872221128791 (2022)

    Bhattacharya, I., Khandwala, Y.S., Vesal, S., Shao, W., Yang, Q., Soerensen, S.J., Fan, R.E., Ghanouni, P., Kunder, C.A., Brooks, J.D., et al.: A review of artificial intelligence in prostate cancer detection on imaging. Therapeutic advances in urology 14, 17562872221128791 (2022)

  8. [8]

    Communications Medicine 2(1), 133 (2022)

    Koh, D.-M., Papanikolaou, N., Bick, U., Illing, R., Kahn Jr, C.E., Kalpathi- Cramer, J., Matos, C., Mart ´ ı-Bonmat ´ ı, L., Miles, A., Mun, S.K.,et al.: Artificial intelligence and machine learning in cancer imaging. Communications Medicine 2(1), 133 (2022)

Show all 84 references
  1. [9]

    Current directions in psychological science 24(4), 267–272 (2015)

    S¨ orqvist, P., Marsh, J.E.: How concentration shields against distraction. Current directions in psychological science 24(4), 267–272 (2015)

  2. [10]

    PloS one 14(11), 0224365 (2019)

    Vabalas, A., Gowen, E., Poliakoff, E., Casson, A.J.: Machine learning algorithm validation with a limited sample size. PloS one 14(11), 0224365 (2019)

  3. [11]

    ACM computing surveys (csur) 53(3), 1–34 (2020)

    Wang, Y., Yao, Q., Kwok, J.T., Ni, L.M.: Generalizing from a few examples: A survey on few-shot learning. ACM computing surveys (csur) 53(3), 1–34 (2020)

  4. [12]

    In: International Conference on Learning Representations (2016)

    Ravi, S., Larochelle, H.: Optimization as a model for few-shot learning. In: International Conference on Learning Representations (2016)

  5. [13]

    Advances in neural information processing systems 30 (2017)

    Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. Advances in neural information processing systems 30 (2017)

  6. [14]

    In: Proceedings 32 of the IEEE/CVF International Conference on Computer Vision, pp

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.-Y., et al.: Segment anything. In: Proceedings 32 of the IEEE/CVF International Conference on Computer Vision, pp. 4015–4026 (2023)

  7. [15]

    Nature Communications 15(1), 654 (2024)

    Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (2024)

  8. [16]

    Nature 622(7981), 156–163 (2023)

    Zhou, Y., Chia, M.A., Wagner, S.K., Ayhan, M.S., Williamson, D.J., Struyven, R.R., Liu, T., Xu, M., Lozano, M.G., Woodward-Court, P., et al.: A foundation model for generalizable disease detection from retinal images. Nature 622(7981), 156–163 (2023)

  9. [17]

    Nature 616(7956), 259–265 (2023)

    Moor, M., Banerjee, O., Abad, Z.S.H., Krumholz, H.M., Leskovec, J., Topol, E.J., Rajpurkar, P.: Foundation models for generalist medical artificial intelligence. Nature 616(7956), 259–265 (2023)

  10. [18]

    Medical Image Analysis 89, 102918 (2023)

    Mazurowski, M.A., Dong, H., Gu, H., Yang, J., Konz, N., Zhang, Y.: Segment anything model for medical image analysis: an experimental study. Medical Image Analysis 89, 102918 (2023)

  11. [19]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp

    Xie, W., Willems, N., Patil, S., Li, Y., Kumar, M.: Sam fewshot finetuning for anatomical segmentation in medical images. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 3253–3261 (2024)

  12. [20]

    In: AAAI 2024 Spring Symposium on Clinical Foundation Models (2024)

    Li, K., Rajpurkar, P.: Adapting segment anything models to medical imaging via fine-tuning without domain pretraining. In: AAAI 2024 Spring Symposium on Clinical Foundation Models (2024)

  13. [21]

    Advances in neural information processing systems 27 (2014)

    Guo, X., Singh, S., Lee, H., Lewis, R.L., Wang, X.: Deep learning for real-time atari game play using offline monte-carlo tree search planning. Advances in neural information processing systems 27 (2014)

  14. [22]

    nature 529(7587), 484–489 (2016)

    Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., et al.: Mastering the game of go with deep neural networks and tree search. nature 529(7587), 484–489 (2016)

  15. [23]

    arXiv preprint arXiv:2408.03314 (2024)

    Snell, C., Lee, J., Xu, K., Kumar, A.: Scaling llm test-time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314 (2024)

  16. [24]

    arXiv preprint arXiv:2402.06457 (2024)

    Hosseini, A., Yuan, X., Malkin, N., Courville, A., Sordoni, A., Agarwal, R.: V- star: Training verifiers for self-taught reasoners. arXiv preprint arXiv:2402.06457 (2024)

  17. [25]

    arXiv preprint arXiv:2309.17179 (2023) 33

    Feng, X., Wan, Z., Wen, M., McAleer, S.M., Wen, Y., Zhang, W., Wang, J.: Alphazero-like tree-search can guide large language model decoding and training. arXiv preprint arXiv:2309.17179 (2023) 33

  18. [26]

    Nature 625(7995), 476–482 (2024)

    Trinh, T.H., Wu, Y., Le, Q.V., He, H., Luong, T.: Solving olympiad geometry without human demonstrations. Nature 625(7995), 476–482 (2024)

  19. [27]

    5: Harnessing proof assistant feedback for reinforcement learning and monte-carlo tree search

    Xin, H., Ren, Z., Song, J., Shao, Z., Zhao, W., Wang, H., Liu, B., Zhang, L., Lu, X., Du, Q., et al.: Deepseek-prover-v1. 5: Harnessing proof assistant feedback for reinforcement learning and monte-carlo tree search. arXiv preprint arXiv:2408.08152 (2024)

  20. [28]

    arXiv preprint arXiv:2501.12948 (2025)

    Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al.: Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 (2025)

  21. [29]

    arXiv preprint arXiv:2409.12917 (2024)

    Kumar, A., Zhuang, V., Agarwal, R., Su, Y., Co-Reyes, J.D., Singh, A., Baumli, K., Iqbal, S., Bishop, C., Roelofs, R., et al.: Training language models to self- correct via reinforcement learning. arXiv preprint arXiv:2409.12917 (2024)

  22. [30]

    arXiv preprint arXiv:2312.08935 (2023)

    Wang, P., Li, L., Shao, Z., Xu, R., Dai, D., Li, Y., Chen, D., Wu, Y., Sui, Z.: Math- shepherd: A label-free step-by-step verifier for llms in mathematical reasoning. arXiv preprint arXiv:2312.08935 (2023)

  23. [31]

    arXiv preprint arXiv:2502.06703 (2025)

    Liu, R., Gao, J., Zhao, J., Zhang, K., Li, X., Qi, B., Ouyang, W., Zhou, B.: Can 1b llm surpass 405b llm? rethinking compute-optimal test-time scaling. arXiv preprint arXiv:2502.06703 (2025)

  24. [32]

    arXiv preprint arXiv:2305.20050 (2023)

    Lightman, H., Kosaraju, V., Burda, Y., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., Cobbe, K.: Let’s verify step by step. arXiv preprint arXiv:2305.20050 (2023)

  25. [33]

    arXiv preprint arXiv:2211.14275 (2022)

    Uesato, J., Kushman, N., Kumar, R., Song, F., Siegel, N., Wang, L., Creswell, A., Irving, G., Higgins, I.: Solving math word problems with process-and outcome- based feedback. arXiv preprint arXiv:2211.14275 (2022)

  26. [34]

    arXiv preprint arXiv:2311.11829 (2023)

    Weston, J., Sukhbaatar, S.: System 2 attention (is something you might need too). arXiv preprint arXiv:2311.11829 (2023)

  27. [35]

    MIT press, ??? (2022)

    Cormen, T.H., Leiserson, C.E., Rivest, R.L., Stein, C.: Introduction to Algo- rithms. MIT press, ??? (2022)

  28. [36]

    Advances in neural information processing systems 35, 24824–24837 (2022)

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V., Zhou, D., et al.: Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35, 24824–24837 (2022)

  29. [37]

    In: Machine Learning Proceedings 1994, pp

    John, G.H., Kohavi, R., Pfleger, K.: Irrelevant features and the subset selec- tion problem. In: Machine Learning Proceedings 1994, pp. 121–129. Elsevier, ??? (1994)

  30. [38]

    ITAT (2019)

    ˇSuch, O., Kontˇ sek, M., Tinajov´ a, A.: Neural pairwise classification models created 34 by ignoring irrelevant alternatives. ITAT (2019)

  31. [39]

    arXiv preprint arXiv:2111.05684 (2021)

    Laakom, F., Chumachenko, K., Raitoharju, J., Iosifidis, A., Gabbouj, M.: Learn- ing to ignore: rethinking attention in cnns. arXiv preprint arXiv:2111.05684 (2021)

  32. [40]

    Advances in Neural Information Processing Systems (2017)

    Vaswani, A.: Attention is all you need. Advances in Neural Information Processing Systems (2017)

  33. [41]

    https://arxiv.org/abs/2410.09918

    Su, D., Sukhbaatar, S., Rabbat, M., Tian, Y., Zheng, Q.: Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces (2024). https://arxiv.org/abs/2410.09918

  34. [42]

    https://arxiv.org/abs/2409

    You, S., Adap, S., Thakur, S., Baheti, B., Bakas, S.: Biochemical Prostate Cancer Recurrence Prediction: Thinking Fast & Slow (2024). https://arxiv.org/abs/2409. 02284

  35. [43]

    Communications of the ACM 63(11), 139–144 (2020)

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks. Communications of the ACM 63(11), 139–144 (2020)

  36. [44]

    Frontiers in psychology 4, 917 (2013)

    Yun, K., Peng, Y., Samaras, D., Zelinsky, G.J., Berg, T.L.: Exploring the role of gaze behavior and object detection in scene understanding. Frontiers in psychology 4, 917 (2013)

  37. [45]

    arXiv preprint arXiv:1803.02999 (2018)

    Nichol, A., Achiam, J., Schulman, J.: On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999 (2018)

  38. [46]

    Science 362(6419), 1140–1144 (2018)

    Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A., Lanc- tot, M., Sifre, L., Kumaran, D., Graepel, T., et al.: A general reinforcement learning algorithm that masters chess, shogi, and go through self-play. Science 362(6419), 1140–1144 (2018)

  39. [47]

    IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)

    Saeed, S.U., Huang, S., Ramalhinho, J., Gayo, I.J., Montana-Brown, N., Bonmati, E., Pereira, S.P., Davidson, B., Barratt, D.C., Clarkson, M.J., et al.: Competing for pixels: a self-play algorithm for weakly-supervised semantic segmentation. IEEE Transactions on Pattern Analysi...

  40. [48]

    In: Medical Image Computing and Computer-assisted intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pp

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical Image Computing and Computer-assisted intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, p...

  41. [49]

    IEEE transactions on pattern analysis and machine intelligence 45(6), 7457–7476 (2022) 35

    Gao, S., Li, Z.-Y., Yang, M.-H., Cheng, M.-M., Han, J., Torr, P.: Large-scale unsupervised semantic segmentation. IEEE transactions on pattern analysis and machine intelligence 45(6), 7457–7476 (2022) 35

  42. [50]

    arXiv preprint arXiv:2210.11016 (2022)

    Gao, S., Zhou, P., Cheng, M.-M., Yan, S.: Towards sustainable self-supervised learning. arXiv preprint arXiv:2210.11016 (2022)

  43. [51]

    The Lancet 389(10071), 815–822 (2017)

    Ahmed, H.U., Bosaily, A.E.-S., Brown, L.C., Gabe, R., Kaplan, R., Parmar, M.K., Collaco-Moraes, Y., Ward, K., Hindley, R.G., Freeman, A., et al.: Diagnostic accuracy of multi-parametric mri and trus biopsy in prostate cancer (promis): a paired validating confirmatory study. Th...

  44. [52]

    Journal of Clinical Oncology 36(36), 3582–3590 (2018)

    Krajewski, K.M., Pedrosa, I.: Imaging advances in the management of kidney cancer. Journal of Clinical Oncology 36(36), 3582–3590 (2018)

  45. [53]

    Cancer imaging 4(Spec No A), 42 (2004)

    Oliva, M.R., Saini, S.: Liver cancer imaging: role of ct, mri, us and pet. Cancer imaging 4(Spec No A), 42 (2004)

  46. [54]

    Cancers 14(6), 1524 (2022)

    Hunter, B., Hindocha, S., Lee, R.W.: The role of artificial intelligence in early cancer diagnosis. Cancers 14(6), 1524 (2022)

  47. [55]

    Annals of internal medicine 165(7), 457–464 (2016)

    Sprague, B.L., Conant, E.F., Onega, T., Garcia, M.P., Beaber, E.F., Herschorn, S.D., Lehman, C.D., Tosteson, A.N., Lacson, R., Schnall, M.D.,et al.: Variation in mammographic breast density assessments among radiologists in clinical practice: a multicenter observational study....

  48. [56]

    Gut 69(6), 977–980 (2020)

    Vleugels, J.L., Koens, L., Dijkgraaf, M.G., Houwen, B., Hazewinkel, Y., Fockens, P., Dekker, E.: Suboptimal endoscopic cancer recognition in colorectal lesions in a national bowel screening programme. Gut 69(6), 977–980 (2020)

  49. [57]

    715–726 (2021)

    Czolbe, S., Arnavaz, K., Krause, O., Feragen, A.: Is segmentation uncertainty useful? In: Information Processing in Medical Imaging: 27th International Con- ference, IPMI 2021, Virtual Event, June 28–June 30, 2021, Proceedings 27, pp. 715–726 (2021). Springer

  50. [58]

    Medical Image Analysis 95, 103181 (2024)

    Saeed, S.U., Ramalhinho, J., Pinnock, M., Shen, Z., Fu, Y., Monta˜ na-Brown, N., Bonmati, E., Barratt, D.C., Pereira, S.P., Davidson, B., et al.: Active learning using adaptable task-based prioritisation. Medical Image Analysis 95, 103181 (2024)

  51. [59]

    arXiv preprint arXiv:2202.09798 (2022)

    Saeed, S.U., Yan, W., Fu, Y., Giganti, F., Yang, Q., Baum, Z., Rusu, M., Fan, R.E., Sonn, G.A., Emberton, M., et al.: Image quality assessment by overlapping task-specific and task-agnostic measures: application to prostate multiparametric mr images for cancer segmentation. ar...

  52. [60]

    Medical Image Analysis 78, 102427 (2022) 36

    Saeed, S.U., Fu, Y., Stavrinides, V., Baum, Z.M., Yang, Q., Rusu, M., Fan, R.E., Sonn, G.A., Noble, J.A., Barratt, D.C., et al.: Image quality assessment for machine learning tasks using meta-reinforcement learning. Medical Image Analysis 78, 102427 (2022) 36

  53. [61]

    In: International Workshop on Applications of Medical AI, pp

    Yan, W., Yang, Q., Syer, T., Min, Z., Punwani, S., Emberton, M., Barratt, D., Chiu, B., Hu, Y.: The impact of using voxel-level segmentation metrics on evaluating multifocal prostate cancer localisation. In: International Workshop on Applications of Medical AI, pp. 128–138 (20...

  54. [62]

    arXiv preprint arXiv:2402.13778 (2024)

    Pocius, M., Yan, W., Barratt, D.C., Emberton, M., Clarkson, M.J., Hu, Y., Saeed, S.U.: Weakly supervised localisation of prostate cancer using reinforcement learning for bi-parametric mr images. arXiv preprint arXiv:2402.13778 (2024)

  55. [63]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Liu, J., Zhang, Y., Chen, J.-N., Xiao, J., Lu, Y., A Landman, B., Yuan, Y., Yuille, A., Tang, Y., Zhou, Z.: Clip-driven universal model for organ segmentation and tumor detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 21152–21164 (2023)

  56. [64]

    In: International Challenge on Kidney and Kidney Tumor Segmentation, pp

    Myronenko, A., Yang, D., He, Y., Xu, D.: Automated 3d segmentation of kidneys and tumors in miccai kits 2023 challenge. In: International Challenge on Kidney and Kidney Tumor Segmentation, pp. 1–7. Springer, ??? (2023)

  57. [65]

    arXiv preprint arXiv:1904.00445 (2019)

    Heller, N., Sathianathen, N., Kalapara, A., Walczak, E., Moore, K., Kaluzniak, H., Rosenberg, J., Blake, P., Rengel, Z., Oestreich, M., et al.: The kits19 challenge data: 300 kidney tumor cases with clinical context, ct semantic segmentations, and surgical outcomes. arXiv prep...

  58. [66]

    Cancers 14(15), 3648 (2022)

    Veiga-Canuto, D., Cerd` a-Alberich, L., Sang¨ uesa Nebot, C., Heras, B., P¨ otschger, U., Gabelloni, M., Carot Sierra, J.M., Taschner-Mandl, S., D¨ uster, V., Ca˜ nete, A., et al.: Comparative multicentric evaluation of inter-observer variability in manual and automatic segmen...

  59. [67]

    arXiv preprint arXiv:2304.07193 (2023)

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., et al.: Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193 (2023)

  60. [68]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Butoi, V.I., Ortiz, J.J.G., Ma, T., Sabuncu, M.R., Guttag, J., Dalca, A.V.: Uni- verseg: Universal medical image segmentation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 21438–21451 (2023)

  61. [69]

    In: International Conference on Machine Learning, pp

    Liang, E., Liaw, R., Nishihara, R., Moritz, P., Fox, R., Goldberg, K., Gonzalez, J., Jordan, M., Stoica, I.: Rllib: Abstractions for distributed reinforcement learning. In: International Conference on Machine Learning, pp. 3053–3062 (2018). PMLR

  62. [70]

    arXiv preprint arXiv:1507.04296 (2015) 37

    Nair, A., Srinivasan, P., Blackwell, S., Alcicek, C., Fearon, R., De Maria, A., Panneershelvam, V., Suleyman, M., Beattie, C., Petersen, S., et al.: Massively parallel methods for deep reinforcement learning. arXiv preprint arXiv:1507.04296 (2015) 37

  63. [71]

    In: Proceedings of the 36th ACM Symposium on Parallelism in Algorithms and Architectures, pp

    Kwok, J., Lohstroh, M., Lee, E.A.: Efficient parallel reinforcement learning frame- work using the reactor model. In: Proceedings of the 36th ACM Symposium on Parallelism in Algorithms and Architectures, pp. 41–51 (2024)

  64. [72]

    IEEE signal processing magazine 35(1), 53–65 (2018)

    Creswell, A., White, T., Dumoulin, V., Arulkumaran, K., Sengupta, B., Bharath, A.A.: Generative adversarial networks: An overview. IEEE signal processing magazine 35(1), 53–65 (2018)

  65. [73]

    IEEE access 7, 36322–36333 (2019)

    Pan, Z., Yu, W., Yi, X., Khan, A., Yuan, F., Zheng, Y.: Recent progress on generative adversarial networks (gans): A survey. IEEE access 7, 36322–36333 (2019)

  66. [74]

    ACM Computing Surveys (CSUR) 54(3), 1–42 (2021)

    Saxena, D., Cao, J.: Generative adversarial networks (gans) challenges, solutions, and future directions. ACM Computing Surveys (CSUR) 54(3), 1–42 (2021)

  67. [75]

    arXiv (2021)

    Antonelli, M., Reinke, A., Bakas, S., Farahani, K., AnnetteKopp-Schneider, Land- man, B.A., Litjens, G., Menze, B., Ronneberger, O., Summers, R.M., Ginneken, B., Bilello, M., Bilic, P., Christ, P.F., Do, R.K.G., Gollub, M.J., Heckers, S.H., Huisman, H., Jarnagin, W.R., McHugo,...

  68. [76]

    Online; accessed Jun 2022 (2013)

    Synapse: Multi-Atlas Labeling Beyond the Cranial Vault - Workshop and Chal- lenge. Online; accessed Jun 2022 (2013). https://doi.org/10.7303/syn3193805 . https://www.synapse.org/#!Synapse:syn3193805/wiki/217789

  69. [77]

    Scientific Data 7(1), 1–9 (2020)

    Rister, B., Yi, D., Shivakumar, K., Nobashi, T., Rubin, D.L.: Ct-org, a new dataset for multiple organ segmentation in computed tomography. Scientific Data 7(1), 1–9 (2020)

  70. [78]

    Medical Image Analysis 69, 101950 (2021)

    Kavur, A.E., Gezer, N.S., Barı¸ s, M., Aslan, S., Conze, P.-H., Groza, V., Pham, D.D., Chatterjee, S., Ernst, P., ¨Ozkan, S., et al.: Chaos challenge-combined (ct- mr) healthy abdominal organ segmentation. Medical Image Analysis 69, 101950 (2021)

  71. [79]

    Advances in neural information processing systems 35, 36722–36732 (2022)

    Ji, Y., Bai, H., Ge, C., Yang, J., Zhu, Y., Zhang, R., Li, Z., Zhanng, L., Ma, W., Wan, X., et al.: Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in neural information processing systems 35, 36722–36732 (2022)

  72. [80]

    The Lancet 389(10071), 815–822 (2017)

    Ahmed, H.U., Bosaily, A.E.-S., Brown, L.C., Gabe, R., Kaplan, R., Parmar, M.K., Collaco-Moraes, Y., Ward, K., Hindley, R.G., Freeman, A., et al.: Diagnostic accuracy of multi-parametric mri and trus biopsy in prostate cancer (promis): a 38 paired validating confirmatory study....

  73. [81]

    Medical image analysis 67, 101821 (2021)

    Heller, N., Isensee, F., Maier-Hein, K.H., Hou, X., Xie, C., Li, F., Nan, Y., Mu, G., Lin, Z., Han, M., et al.: The state of the art in kidney and kidney tumor segmentation in contrast-enhanced ct imaging: Results of the kits19 challenge. Medical image analysis 67, 101821 (2021)

  74. [82]

    Nature methods 18(2), 203–211 (2021)

    Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods 18(2), 203–211 (2021)

  75. [83]

    IEEE access 6, 58774–58783 (2018)

    Liang, D., Yang, F., Zhang, T., Yang, P.: Understanding mixup training methods. IEEE access 6, 58774–58783 (2018)

  76. [84]

    arXiv preprint arXiv:1710.09412 (2017) 39

    Zhang, H.: mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017) 39

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.