Glance and Focus Reinforcement for Pan-cancer Screening

AI summary

Overview

Research area: Medical computer vision — automatic detection and segmentation of multiple cancer types (pan-cancer screening) in large 3D CT scans, combining deep learning with reinforcement learning.

Technical level: Advanced. The paper assumes familiarity with 3D medical image segmentation (nnUNet, SwinUNETR), reinforcement learning (policy gradients, PPO, GRPO), and clinical screening metrics (DSC, NSD, F1, sensitivity/specificity).

Scope: The paper introduces GF-Screen, a two-model "glance and focus" framework that uses a lightweight classifier to discard healthy regions of a CT volume and a segmentation network to analyze the rest, with a group-relative reinforcement learning signal tying the two together.

What This Paper Is About

In large CT scans, tumors typically occupy a tiny fraction of the volume — in this study's dataset, just 0.085% of the total imaged area. Existing AI screening models must scan every region at high resolution, which wastes computation on healthy tissue and produces false alarms. This paper proposes a system that mimics how radiologists work: first glance broadly across the scan to spot suspicious areas, then focus computational effort only there. The goal is a single model that can screen for nine different lesion types accurately, quickly, and with few false positives.

Key Contributions

1. A glance-and-focus framework for pan-cancer screening. GF-Screen pairs a lightweight 3D ResNet-18 "Glance" model that classifies cropped sub-volumes as lesion-containing or healthy with an SwinUNETR "Focus" model that performs precise pixel-level segmentation only on the selected sub-volumes.

2. Reinforcement learning driven by segmentation quality. Because the Glance model's selection step is non-differentiable, the authors train it with RL, using the Focus model's segmentation output as the reward signal. A sub-volume earns reward 1 if the predicted mask overlaps the ground-truth lesion at all, and 0 otherwise.

3. A Group Relative Learning (GRL) paradigm for vision. The paper adapts the GRPO technique from large language models to 3D medical vision by exploiting a natural property of the task: the group of cropped sub-volumes already provides comparable candidates, so no language model or extra value network is needed to estimate relative advantages. High-advantage sub-volumes are prioritized; low-advantage ones are discarded.

4. Large-scale validation and a decisive benchmark result. Training and validation used 5,117 CT scans from 23 datasets (16 internal, 7 external) covering 9 lesion types, with GF-Screen placing first on the public MICCAI FLARE25 pan-cancer challenge validation leaderboard.

Main Findings

• Segmentation accuracy: GF-Screen achieves 60.8% average DSC across nine lesion types on internal datasets, beating the second-best automatic method (VoCo, 56.1%) by 4.7 points and outperforming specialized pan-cancer models like PASTA and SuPreM.

• Detection accuracy: 95.9% average F1-score across lesion types, versus 92.5% for SwinUNETR, with perfect detection (100%) on COVID-19 infection, kidney tumor, and adrenocortical carcinoma.

• External generalization: On three unseen datasets (Rider, Corona, IRCADb), GF-Screen reaches 54.1% average DSC, surpassing nnUNet by 5.1 points, showing the gains are not dataset-specific.

• False positive reduction: Average false positive rate of 15.6% on healthy datasets versus 30.4% for nnUNet and 38.7% for SuPreM — roughly 23% lower than SuPreM. The authors attribute this to discarding low-advantage healthy regions rather than forcing the model to evaluate them.

• Efficiency: The Glance model discards 83.3% of sub-volumes while retaining 97.7% sensitivity for lesion-containing ones, reducing computation from 12,155 to 2,164 GFLOP (5.7x reduction) and inference time from 114 seconds to 28 seconds per scan.

• Challenge leaderboard: On the FLARE25 pan-cancer validation set, GF-Screen scores 58.6% DSC and 52.2% NSD, exceeding the FLARE24 champion (33.0% DSC, 24.0% NSD) by +25.6% and +28.2% respectively.

• RL design ablation: Plain cross-entropy classification training collapses (the Glance model selects only 3.1% of sub-volumes, causing DSC to drop to 37.6%). PPO with a value model performs poorly (24.5% DSC). Using DSC as the reward gives 43.2%. The binary detection reward with GRL reaches 53.1%, and adding a small classification loss term brings it to 56.7%.

• Hyperparameter findings: A group size of N=16 works best; larger or smaller values hurt. Glance backbone choice matters little (ResNet-18, ResNet-34, and a 3D UNet all land near 56-57% DSC).

• Backbone generality: Adding GF-Screen to weaker segmentation backbones also helps (UNETR: 16.2% to 27.3% DSC; TransUNet: 27.2% to 34.8%; nnUNet: 43.7% to 45.6%), though the strongest combination is with SwinUNETR.

Methodology in Plain English

The system splits the problem into two stages.

Stage one — Glance. A CT scan is cut into many overlapping sub-volumes of size 96×96×64 using a sliding window. A small classification network looks at each one and outputs a probability that it contains a lesion. Only sub-volumes above the decision threshold get passed downstream.

Stage two — Focus. A segmentation network processes the surviving sub-volumes and produces pixel-level masks outlining lesions.

The training trick is that these two stages are trained jointly but the selection step is a hard yes/no decision, which blocks gradients from flowing back from the segmentation loss into the classifier. The authors solve this with reinforcement learning: the Glance model is a policy that picks actions (select or discard), and the Focus model's output provides the reward.

Rather than relying on a hand-crafted reward or a learned value network, the authors use the group of sub-volumes as a built-in comparison set. They normalize each sub-volume's reward against the mean and standard deviation of its group, producing an "advantage" score. Sub-volumes with high advantage are reinforced; those with low advantage are suppressed. This replaces GRPO's need for an LLM to generate candidate answers — the sub-volume group is the candidate set.

The training objective combines the RL policy-gradient term (with clipping and a KL penalty against a frozen reference model for stability) with a small cross-entropy classification loss to keep the model from collapsing. A key design decision: the reward is binary (did the predicted mask touch the true lesion at all?) rather than DSC-based, because the authors argue a high DSC can just mean the sub-volume was easy, while a low DSC may indicate a hard but clinically important case that should not be discarded. This mirrors prior cancer screening work that deliberately avoids filtering out difficult cases.

At inference, no post-processing is applied; the sliding window plus Glance filtering alone produces the final output.

Why This Matters

Impact on research. This is the first framework to apply modern group-relative reinforcement learning to pan-cancer screening, showing that RL techniques developed for language models can be transferred to dense 3D medical vision tasks without an LLM or a value network. The binary-reward design and the exploitation of sub-volume groups as natural candidate sets are reusable ideas for other imbalanced detection problems.

Real-world applications:

• Routine physical-examination CT screening, where a single model could flag multiple cancer types across chest and abdomen without organ-specific pipelines.
• Opportunistic screening, where scans taken for unrelated reasons could be automatically checked for incidental lesions.
• Resource-constrained clinical deployment, where the 5.7x speedup and 16 GB inference memory footprint make large-scale screening feasible on modest hardware.
• False-alarm reduction in radiology workflows, cutting the number of healthy cases radiologists must manually review.

Industry relevance. The combination of higher accuracy, lower false positives, and a 4x inference speedup directly addresses the cost and trust barriers that block clinical adoption of AI screening tools. Vendors of CT scanners, PACS software, and teleradiology services could integrate such a model to triage scans before human reading.

Future Directions

• Extending to more modalities and lesion types. The current evaluation covers CT only and nine lesion types; whether the glance-and-focus reward structure transfers to MRI, PET, or ultrasound is untested.
• Improving Glance specificity. Sensitivity is 97.7% but specificity is 75.9%, meaning roughly a quarter of healthy sub-volumes still reach the Focus model. Better filtering could increase the speedup further.
• Understanding the reward function more deeply. The authors note that DSC-based rewards underperformed, but the space of possible reward shapes (partial credit, lesion-size weighting, clinically weighted rewards) is largely unexplored.
• Reducing dependence on a frozen reference model. The KL-regularized GRL setup requires maintaining and periodically updating a reference network; whether simpler or more adaptive alternatives work as well is an open question.
• Prospective clinical validation. All evaluation is on retrospective public datasets; the system has not been tested in a live screening workflow where prevalence, scanner variety, and patient population differ.

Target Audience

Medical imaging researchers and graduate students working on 3D segmentation, lesion detection, or computer-aided diagnosis; machine learning researchers interested in applying reinforcement learning beyond language and control; and clinical AI engineers or radiology informatics specialists evaluating whether pan-cancer screening models are ready for deployment. Readers without background in segmentation metrics or policy-gradient methods will find the experimental sections dense, though the high-level glance-and-focus analogy is accessible to anyone familiar with radiology practice.

Authors’ abstract

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).