Glance and Focus Reinforcement for Pan-cancer Screening
Linshan Wu, Jiaxin Zhuang, Hao Chen
TL;DR
GF-Screen is a novel reinforcement learning framework inspired by radiologists' glance and focus strategy, improving pan-cancer screening accuracy by localizing and segmenting tiny lesions in large CT volumes, reducing false positives and increasing efficiency.
Contribution
Introduces GF-Screen, a reinforcement learning-based approach that combines a Glance model and a Focus model for effective pan-cancer screening in CT scans, with a novel group relative learning paradigm.
Findings
Outperforms previous methods on multiple datasets.
Achieves top results on MICCAI FLARE25 challenge leaderboard.
Reduces false positives and improves lesion localization accuracy.
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…
Peer Reviews
Decision·ICLR 2026 Poster
1. The results are the most significant strength. Leading the FLARE25 challenge validation leaderboard and outperforming the FLARE24 champion solution by such a large margin (+25.6% DSC) is a truly remarkable achievement. 2. The method directly addresses a critical bottleneck in deploying AI screening models: computational cost and false positives. By intelligently discarding healthy regions, the 5.7x computational saving is a major practical breakthrough, making large-scale screening far more f
1. The paper introduces GRL as a "novel" paradigm, but its precise mechanism and distinction from existing policy gradient methods (like PPO, which is cited) could be clearer. The loss function in Eq. 3 looks like a PPO-style clipped objective. The novelty seems to be in the advantage estimation or the "group relative comparison", but this specific aspect is not explained in sufficient detail. This lack of clarity is particularly problematic concerning the core RL mechanism. For instance, the pa
Strengths 1. The “glance and focus” paradigm is a clever conceptual and algorithmic adaptation of clinical diagnostic reasoning to AI-based screening. 2. The paper provides a well-defined pipeline linking reinforcement learning and segmentation, with precise mathematical formulations for the GRL optimisation objective. 3. Addressing pan-cancer screening rather than single-organ segmentation is highly ambitious and clinically meaningful. 4. The study presents extensive experiments on both interna
Weaknesses 1. While the proposed framework is well-engineered, its novelty may primarily lie in combining existing ideas (sub-volume selection, segmentation, RL reward optimisation) rather than introducing a fundamentally new theoretical concept. 2. The reinforcement learning formulation, especially the reward design and advantage estimation, is largely empirical. There is little discussion of convergence, stability, or theoretical motivation for using group relative learning in medical imaging.
1. The authors effectively reframe the well-established coarse-to-fine paradigm as a "glance-and-focus" strategy, an evocative and clinically intuitive framing that demonstrates strong scientific storytelling. 2. The figures are clean, well-designed, and complemented by clear, concise textual descriptions, making the methodology easy to follow. 3. The empirical evaluation is extensive, with experiments across multiple internal and external datasets and diverse lesion types, providing thorough su
1. The motivation for using reinforcement learning (RL) to train the Glance model is underdeveloped. While extreme class imbalance is acknowledged as a key challenge, the authors do not sufficiently justify why conventional approaches, such as intelligent sampling, focal loss, or hard-negative mining, would be inadequate. Jumping directly to RL, a currently popular but complex paradigm, feels more like a methodological trend than a necessity. 2. Within the RL framework, using segmentation perfor
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAI in cancer detection · COVID-19 diagnosis using AI · Radiomics and Machine Learning in Medical Imaging