Glance and Focus Reinforcement for Pan-cancer Screening

TL;DR

This work addresses pan-cancer screening in large-scale CT volumes where tiny lesions and extreme foreground-background imbalance hinder existing AI methods. It proposes GF-Screen, a Glance and Focus reinforcement learning framework that uses a light Glance model to crop sub-volumes and a Focus model to segment lesions, with segmentation results providing reward signals to guide Glance training via a group-relative learning scheme. A novel Group Relative Learning paradigm prioritizes high-advantage sub-volumes within groups, improving efficiency and reducing false positives. Experiments across 16 internal and 7 external datasets covering 9 lesion types show GF-Screen achieving state-of-the-art performance, including leading the MICCAI FLARE25 leaderboard with substantial gains in DSC and NSD, while reducing computation by approximately 5.7x. This work demonstrates a practical RL-based approach for dense 3D medical imaging and points toward future clinician-in-the-loop deployment.

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).

Figures (12)

• Figure 1: (a) The pan-cancer dataset used in this study, encompassing 5,117 CT scans across 9 different types of lesions from 16 internal and 7 external datasets. (b) Significant foreground-background imbalance: in our dataset, lesions occupy only 0.085% area proportions in CT volumes. (c) Comparisons in pan-cancer segmentation. (d) Comparisons in pan-cancer detection. (e) Inference efficiency (GFLOP per scan) and segmentation DSC on the FLARE23 validation dataset. Compared with the second-best model, GF-Screen is 5.7$\times$ faster with higher segmentation DSC. (f) Comparisons on the FLARE25 challenge validation leaderboard. GF-Screen outperforms the second-ranked algorithm (champion solution of FLARE24) by a large margin (+25.6% DSC and +28.2% NSD).
• Figure 2: Comparison between previous cancer screening methods and GF-Screen.
• Figure 3: The overall framework of GF-Screen, including a Glance model to localize diseased regions and a Focus model to precisely segment lesions. (a) In the training stage, we conduct segmentation on all sub-volumes and leverage the segmentation results to reward the Glance model via a novel group-relative learning paradigm. (b) In the inference stage, a dynamic number of sub-volumes classified as "with lesions" by the Glance model will be input to the Focus model for segmentation, where the redundant regions will be discarded.
• Figure 4: Illustration of sub-volume variation. The blue regions represent the challenging view with partial lesions and poor angles. While red regions indicate the optimal diagnostic view containing complete lesion information, generally with more precise segmentation results. Thus, we propose to leverage segmentation results as reward signals for RL.
• Figure 5: The Group Relative Learning (GRL) paradigm in GF-Screen. The Glance model $G$ is trainable while the reference model $G_{ref}$ is frozen. We first generate selection outputs $o$ from the input sub-volumes $v$, then use the reward function Eq. \ref{['eqn_reward']} to calculate the rewards $r$. Finally, we compute the relative advantages $A$ via the GAE function Eq. \ref{['eqn_advantage']}.