PaperlayerView exact revision on arXiv

Source-linked AI summary

Glance and Focus Reinforcement for Pan-cancer Screening

Linshan Wu, Jiaxin Zhuang, Hao Chen

arXiv:2601.19103v2cs.CV

TL;DR

Pan-cancer screening in large CT volumes is challenged by tiny, diverse lesions and severe foreground-background imbalance, which makes localization difficult and increases inefficient healthy-region processing. GF-Screen combines Glance localization, Focus segmentation, and reinforcement learning with group-relative selection. Across 16 internal and 7 external datasets covering 9 lesion types, it led the MICCAI FLARE25 validation leaderboard, exceeding the FLARE24 champion by +25.6% DSC and +28.2% NSD.

  • Problem

    Tiny lesions in large CT volumes and extreme foreground-background imbalance make diverse lesion localization difficult while redundant healthy-region focus reduces efficiency and increases false positives.

  • Method

    GF-Screen combines a Glance model that selects lesion-containing sub-volumes, a Focus model that segments lesions, and reinforcement learning that uses Focus segmentation results to reward Glance.

  • Results

    +25.6% DSC and +28.2% NSD over the FLARE24 champion solution were reported on the MICCAI FLARE25 public validation leaderboard.

  • Takeaways & Limitations

    GF-Screen provides a pan-cancer screening framework evaluated across 16 internal and 7 external datasets covering 9 lesion types.

Abstract

from arXiv · show

Pan-cancer screening in large-scale CT scans remains challenging for existing AI methods, primarily due to the difficulty of localizing diverse types of tiny lesions in large CT volumes. The extreme foreground-background imbalance significantly hinders models from focusing on diseased regions, while redundant focus on healthy regions not only decreases the efficiency but also increases false positives. Inspired by radiologists' glance and focus diagnostic strategy, we introduce GF-Screen, a Glance and Focus reinforcement learning framework for pan-cancer screening. GF-Screen employs a Glance model to localize the diseased regions and a Focus model to precisely segment the lesions, where segmentation results of the Focus model are leveraged to reward the Glance model via Reinforcement Learning (RL). Specifically, the Glance model crops a group of sub-volumes from the entire CT volume and learns to select the sub-volumes with lesions for the Focus model to segment. Given that the selecting operation is non-differentiable for segmentation training, we propose to employ the segmentation results to reward the Glance model. To optimize the Glance model, we introduce a novel group relative learning paradigm, which employs group relative comparison to prioritize high-advantage predictions and discard low-advantage predictions within sub-volume groups, not only improving efficiency but also reducing false positives. In this way, for the first time, we effectively extend cutting-edge RL techniques to tackle the specific challenges in pan-cancer screening. Extensive experiments on 16 internal and 7 external datasets across 9 lesion types demonstrated the effectiveness of GF-Screen. Notably, GF-Screen leads the public validation leaderboard of MICCAI FLARE25 pan-cancer challenge, surpassing the FLARE24 champion solution by a large margin (+25.6% DSC and +28.2% NSD).

1 INTRODUCTION

Pan-cancer screening seeks one model for diverse lesions in large CT volumes, but tiny lesions and severe foreground-background imbalance make localization difficult and healthy-region focus inefficient. GF-Screen addresses this with radiologist-inspired Glance and Focus reinforcement learning and reports strong multi-dataset and challenge results.

  • Pan-cancer screening aims to detect and segment different lesion types with one universal model in large-scale CT scans.CT is commonly used in routine physical examinations, giving pan-cancer screening clinical significance.
  • GF-Screen was evaluated on 5,117 CT scans spanning 16 internal and 7 external datasets and 9 lesion types.
  • Lesions occupy tiny areas in large CT volumes, creating extreme foreground-background imbalance that hinders localization and increases false positives from healthy-region focus.
  • GF-Screen combines a Glance model for coarse diseased-region localization with a Focus model for precise lesion segmentation, using Focus segmentation results to reward Glance through reinforcement learning.The Glance model selects lesion-containing sub-volumes for Focus segmentation because selection is non-differentiable for segmentation training.
  • +25.6% DSC and +28.2% NSD over the FLARE24 champion solution were achieved on the MICCAI FLARE25 public validation leaderboard.GF-Screen led the reported validation leaderboard.

2 RELATED WORKS

Prior pan-cancer methods often specialize in individual lesion types and use sliding-window inference that retains redundant healthy regions. GF-Screen instead combines coarse localization, precise segmentation, and reinforcement learning to select relevant sub-volumes.

  • 2.1 CANCER SCREENING: Existing cancer-screening approaches primarily use segmentation models but are often specialized to a single lesion type.
  • 2.1 CANCER SCREENING: Prior methods generally apply sliding-window inference to large CT volumes without discarding redundant healthy regions.
  • 2.1 CANCER SCREENING: Retaining healthy regions impedes inference efficiency and increases false positives on healthy regions.
  • 2.1 CANCER SCREENING: GF-Screen uses a Glance model for coarse localization and a Focus model for pixel-level lesion segmentation.The Glance model classifies cropped sub-volumes as with or without lesions.
  • 2.2 REINFORCEMENT LEARNING IN VISION PERCEPTION: GF-Screen applies reinforcement learning to reward Glance using Focus segmentation results, with selected lesion-containing sub-volumes passed to Focus while redundant regions are discarded.

3 METHOD

GF-Screen combines a Glance model for coarse sub-volume localization with a Focus model for precise lesion segmentation. Reinforcement learning uses Focus segmentation outcomes to train Glance to prioritize diagnostically useful views and discard redundant regions.

  • 3.1 TOWARDS PRECISE AND EFFICIENT PAN-CANCER SCREENING: GF-Screen uses a lightweight Glance classifier to identify lesion-containing sub-volumes and a Focus model to precisely segment lesions.The two models can integrate advanced network architectures.
  • 3.2 GLANCE AND FOCUS REINFORCEMENT LEARNING: The framework trains Focus with lesion masks while training Glance through reinforcement-learning rewards derived from Focus segmentation results.Glance selection is non-differentiable within Focus segmentation training, motivating reinforcement learning.
  • 3.2 GLANCE AND FOCUS REINFORCEMENT LEARNING: Random cropping creates lesion-containing views with partial lesions or poor angles, so binary lesion labels alone can hamper Glance classification.Severe foreground-background imbalance also encourages overfitting to negative sub-volumes and reduces positive sensitivity.
  • 3.2 GLANCE AND FOCUS REINFORCEMENT LEARNING: The Focus model supplies a binary reward of 1 when its predicted segmentation overlaps the lesion mask and 0 otherwise.The authors report that using segmentation DSC as a more granular reward performed worse because DSC varies with lesion complexity and view difficulty.
  • 3.3 GROUP RELATIVE LEARNING FOR SUB-VOLUMES SELECTION: Group Relative Learning compares rewards among cropped sub-volumes, prioritizing high-advantage selections and discarding low-advantage predictions without an additional candidate-generation mechanism.Group reward normalization derives relative advantages, while the objective also includes KL regularization and a small classification-loss coefficient.

4 EXPERIMENTS

GF-Screen is evaluated across internal and external datasets, lesion types, efficiency, challenge benchmarks, and reinforcement-learning ablations. Results show strong segmentation and detection performance while reducing redundant computation and false positives.

  • External validation: 54.1% average DSC on three external datasets surpasses nnUNet by 5.1%.The external datasets cover lung tumors, COVID-19, and liver tumors.
  • False positives: 23.1% lower false positives than SuPreM demonstrates improved behavior on healthy datasets.The authors associate competing models’ higher false positives with redundant focus on healthy regions.
  • Challenge performance: GF-Screen achieves 56.7% DSC on FLARE23 and surpasses previous methods by a clear margin.Inference duration is also evaluated because efficiency matters for large-scale screening deployment.
  • Challenge performance: +25.6% DSC and +28.2% NSD over the FLARE24 champion solution establish GF-Screen as the leading method on the FLARE25 validation leaderboard.The comparison is reported under the FLARE25 validation leaderboard evaluation.
  • Efficiency and selection: 97.7% sensitivity is achieved while discarding 83.3% of redundant sub-volumes and reducing computation cost by 5.7 times.The preserved sub-volume ratio is 16.7%.

5 CONCLUSION

The paper presents GF-Screen as a Glance and Focus reinforcement-learning framework for pan-cancer screening, using segmentation feedback to guide region selection.

  • Conclusion: GF-Screen selects diseased regions for Focus-model segmentation and discards redundant healthy regions.The framework uses Focus-model segmentation results as rewards for the Glance model through reinforcement learning.
  • Conclusion: Experiments across 16 internal and 7 external datasets and 9 lesion types demonstrate superior performance.The authors describe the framework as a practical conceptual contribution rather than a new RL theory.

A. Emre Kavur, N. Sinem Gezer, Mustafa Barıs¸, Sinem Aslan, Pierre-Henri Conze, Vladimir Groza,

This appendix passage set consists of bibliographic entries cited throughout the paper.

  • References: Several entries document datasets and cancer-screening applications used as related evidence or evaluation resources.Examples include pancreas CT, IRCAD liver segmentation, and breast cancer screening.
  • References: The bibliography also includes reinforcement-learning and visual-search literature relevant to the framework.Referenced works include Proximal Policy Optimization and adaptive visual exploration.

6 APPENDIX

The appendix documents datasets, metadata, comparison methods, evaluation procedures, and preprocessing or training settings.

  • Datasets: The study uses 5,117 CT scans from 16 internal and 7 external datasets covering 9 lesion types.The Atlas dataset’s healthy cases were selected to avoid overlap with other datasets.
  • Datasets: Dataset metadata include scanner, contrast phase, and voxel size, although some public datasets lack device or contrast information.The metadata are collected from public dataset links.
  • Comparison methods: Comparison methods include nnUNet, SwinUNETR, 3D UX-Net, CLIP-driven, TransUNet, UniMiss+, VoCo, SuPreM, and PASTA.External comparisons are restricted when those datasets were already included in comparison-model training.
  • Evaluation: Segmentation is evaluated with DSC, while mask-overlap detection is reported using scan-level F1-Score and false-positive rates.The detection protocol treats a scan as detected when the segmentation prediction overlaps the ground truth.
  • Preprocessing and training: Appendix tables provide preprocessing details and training settings for the experiments.These settings are summarized in Table A4.

6.2 SUPPLEMENTARY RESULTS

Supplementary analyses examine GF-Screen across datasets, repeated runs, sliding-window settings, efficiency, and visual detection behavior. The results show faster inference and coarse lesion localization, while also documenting missed tiny lesions as failure cases.

  • Dataset-wise evaluation: Pan-cancer segmentation performance is reported on 16 datasets across 9 lesion types using DSC (%).The supplementary table reports results for automatic screening models and includes no-lesion cases selected from Atlas.
  • Stability analysis: Five GF-Screen experiments report mean and standard deviation for segmentation DSC and detection F1-Score across 9 lesion types.Tables A6 and A7 quantify training-run variation for both segmentation and detection.
  • Ablation and significance: Ablation studies evaluate sliding-window sizes across different lesion types, while t-test p-values compare GF-Screen with VoCo.These analyses assess sensitivity to window size and statistical differences from the best-competing method.
  • Efficiency and comparison: 5.7 times faster with higher performance simultaneously than the second-best model, according to the supplementary comparison.The comparison is summarized in Fig. A1.
  • Glance-model analysis: The Glance model detects lesions at a coarse sub-volume level and provides diagnostic views for Focus segmentation, but can miss tiny lesions or indicate imprecise regions.Visual activation maps are compared with ground-truth lesion masks, including failure cases.

6.4 REDUCE FALSE POSITIVES

GF-Screen reduces false positives and improves segmentation by discarding redundant healthy or suboptimal views during inference. Its supplementary visualizations illustrate lesion-type results, filtered views, and reinforcement-learning loss curves.

  • False-positive reduction: 23.1% fewer false positives on healthy regions improves overall DSC scores.GF-Screen discards redundant regions during inference to address false positives reported for previous methods.
  • Optimal-view selection: Selecting optimal diagnostic views improves segmentation DSC by retaining complete-organ views and filtering challenging partial-information views.Previous methods average overlapping sliding-window predictions, allowing challenging views to degrade final results.
  • Optimal-view selection: Figure A4 contrasts unfiltered predictions, whose challenging views can degrade average performance, with GF-Screen’s retention of optimal-view results.The figure presents an illustrative case of view selection.
  • Visual evaluation: Segmentation results are presented across different lesion types, and healthy-region segmentation results are compared with Wu et al. (2025c).These visual comparisons complement the false-positive analysis.
  • Training analysis: Training loss curves are analyzed by randomly splitting the training dataset into five subsets and retraining.Figure A6 uses the resulting curves to analyze reinforcement-learning training behavior.

6.7 DERIVATION OF THE KL DIVERGENCE APPROXIMATION

The KL regularization used in Eq. 4 is derived as a second-order Taylor approximation of reverse KL divergence. The resulting surrogate preserves non-negativity around matching policies while being computationally friendlier for GRPO-style optimization.

  • Derivation: The regularization term D_KL in Eq. 4 originates from GRPO and approximates standard reverse KL divergence through a second-order Taylor expansion.The derivation expands the relevant function around the point where G and G_ref are identical.
  • Surrogate properties: The surrogate g(r) = r − log r − 1 is non-negative and reaches its minimum of 0 at r = 1.Its expectation under G recovers the form used in the objective.
  • Optimization role: The ratio-based surrogate penalizes deviations of policy G from reference policy G_ref and supports stable, easier-to-optimize policy training.The passage motivates this form as a computationally friendly alternative to exact KL divergence in GRPO.

6.8 THE USE OF LARGE LANGUAGE MODELS (LLMS)

The authors use DeepSeek LLMs only for grammar-error checking, not for the research method or analysis.

  • LLM use: DeepSeek LLMs are used only to check for grammar errors.No other use is specified.

6.9 REPRODUCIBILITY STATEMENT

GF-Screen reports publicly validated code and checkpoints while identifying broader algorithmic, dataset, and clinical evaluation work as future directions. Its effectiveness in clinical practice remains to be established on real-world datasets.

  • The authors promise to release GF-Screen’s code and model checkpoints, with Docker-based public leaderboard validation.
  • The framework is presented as the first reinforcement learning approach for pan-cancer screening and could incorporate policy optimization algorithms beyond GRPO.
  • Future work will expand datasets and lesion types, explore head-neck CT lymph-node detection, and apply group relative learning to other medical image analysis tasks.
  • Clinical effectiveness remains unsubstantiated because evaluation has so far used public datasets rather than real-world clinical data.
Loading 2601.19103v2…