💻 computer science

Interactive Whole Slide Images for RL-based Tumour Segmentation

This paper proposes an end-to-end reinforcement learning framework that treats whole-slide images as hierarchical multi-resolution environments, enabling an agent to perform sequential tumour segmentation via navigation and zooming actions, which achieves comparable coarse segmentation quality to patch-based methods while significantly reducing inference time.

Original authors: Mohamad Mohamad, Francesco Ponzio, Maxime Gassier, Nicolas Pote, Xavier Descombes

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohamad Mohamad, Francesco Ponzio, Maxime Gassier, Nicolas Pote, Xavier Descombes

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of medical diagnosis, pathologists examine tissue samples under microscopes to identify diseases like cancer. To do this digitally, they use whole-slide images, which are massive, high-resolution scans of entire glass slides. These images are so large that they contain billions of pixels, far more than a computer screen can display at once. To make sense of them, the images are stored as pyramids of layers, where the top layer is a tiny, blurry overview and the bottom layers are incredibly detailed close-ups. Traditionally, computers analyze these slides by chopping them into thousands of small, fixed squares and checking each one individually. While this works, it is slow and wasteful, often forcing the computer to examine vast areas of healthy tissue just to find the small, scattered regions where a tumor might hide.

Researchers have long wondered if computers could learn to look at these slides the way a human doctor does: by scanning the overview, zooming in on suspicious spots, and ignoring the rest. A new study by a team from France, Italy, and Paris proposes a method where an artificial intelligence agent learns to navigate these digital slides on its own. Instead of processing the entire image at once, the agent moves through the slide, deciding where to look, how much to zoom, and when to mark a spot as a tumor. The researchers found that this interactive approach allows the computer to find tumors with accuracy similar to traditional methods but in a fraction of the time, turning a process that takes minutes into one that takes seconds.

The core of this work is a system that treats the digital slide not as a static picture, but as an interactive environment. Imagine a pathologist sitting at a microscope; they do not look at every single cell in the sample. Instead, they start with a broad view, spot an area that looks unusual, and then zoom in to inspect it closely. The researchers built a computer program that mimics this behavior. They created a virtual agent that can move left, right, up, or down, and can zoom in or out to change the level of detail it sees. The agent's goal is to build a map of the tumor by making a series of these moves and selections. It starts with a blank slate and gradually fills in the tumor's shape by choosing which parts of the slide to focus on.

To teach the agent how to do this, the researchers used a learning method called reinforcement learning. This is a type of training where the agent learns by trial and error, receiving feedback for its actions. In this case, the agent gets a reward when its current guess about the tumor's location gets closer to the true location, which is known from expert human annotations. The researchers designed the reward system carefully. They found that if the agent is rewarded simply for finding any part of the tumor, it might become too aggressive and mark healthy tissue as cancerous. However, if the reward is structured to encourage the agent to refine its map as it goes, the agent learns to be more precise. The study showed that giving the agent a moderate amount of time to make its moves—around 80 steps per slide—allowed it to find the tumor effectively without getting lost in unnecessary details.

The team tested this system on slides from patients with lung adenocarcinoma, a common type of lung cancer. They compared their new agent against standard computer methods that chop the slide into fixed squares. The results were striking. The traditional methods, which process the entire slide, took an average of over 229 seconds to analyze a single slide, with some taking much longer depending on the size of the tumor. In contrast, the new interactive agent completed the same task in an average of just 3 seconds. While the agent was slightly less precise at drawing the exact edges of the tumor compared to the most advanced traditional methods, it captured the overall shape and location of the disease with comparable accuracy. This trade-off is significant because it means a doctor could get a preliminary analysis of a slide almost instantly, rather than waiting for a computer to process every inch of the image.

One of the most important discoveries in the study was that the agent needed to see a wide variety of examples to learn well. When the researchers trained the agent on a small set of slides, it struggled to generalize to new patients. However, when they expanded the training set to include many more slides, the agent became much more reliable. This suggests that for computers to learn how to navigate complex medical images effectively, they need to be exposed to a diverse range of cases, just as human doctors do. The study also highlighted that the agent did not need complex memory systems to remember what it had seen; the combination of the current view and a global overview of the slide was enough for it to make good decisions.

The researchers acknowledge that their method is not perfect. The agent sometimes struggles with the fine details of the tumor's boundary, likely because it moves in steps that are too large to catch the very smallest irregularities. Additionally, the study was conducted on a single type of cancer from one hospital, so it is not yet known if the same approach will work for other diseases or in different clinical settings. Despite these limitations, the work demonstrates a clear path forward. By allowing computers to interact with medical images in a way that mirrors human exploration, it is possible to drastically reduce the time required for diagnosis without sacrificing too much accuracy. This approach offers a promising way to handle the overwhelming volume of data in modern pathology, potentially allowing doctors to focus their attention on the cases that need it most.

Technical Summary: Interactive Whole Slide Images for RL-based Tumour Segmentation

Problem Statement
Whole-slide images (WSIs) present significant computational challenges for tumour segmentation due to their extreme spatial resolution and the sparse distribution of tumour regions. Traditional deep learning approaches typically decompose WSIs into predefined collections of image patches, processing them independently or aggregating results. While effective, these methods often discard the sequential and interactive nature of slide exploration, leading to the exhaustive processing of redundant or diagnostically irrelevant regions. Furthermore, existing Reinforcement Learning (RL) applications in histopathology have largely focused on selecting informative patches from predefined candidate sets or constrained tile spaces, rather than treating the hierarchical WSI pyramid itself as an interactive environment. This limits the ability to model the coarse-to-fine navigation strategies employed by pathologists.

Methodology
The authors propose an end-to-end reinforcement learning framework that formulates the WSI pyramid as a hierarchical, multi-resolution interactive environment. The task is defined as a Markov Decision Process (MDP) where an agent navigates the slide to sequentially construct a tumour segmentation mask.

Environment Design: The WSI is modeled as a spatial multi-resolution pyramid (levels $0$ to $n$ ) plus a global thumbnail. The agent navigates this environment using movement (spatial translation) and zooming (level transitions) actions. The state representation ( $s_t$ ) at each timestep consists of a pair of observations: a current local patch ( $o_{current}$ ) and a global thumbnail ( $o_{global}$ ), both augmented with an accumulated selection mask from previous steps.
Agent Architecture: The agent utilizes an actor-critic architecture with a shared backbone encoder (ResNet-family) to process local and global observations. The features are fused and passed to separate actor (policy) and critic (value) heads. The agent is trained using Proximal Policy Optimization (PPO).
Action Space: Actions are divided into navigation (spatial movement and zooming) and decision (tumour selection). The study investigates two configurations: a 7-action space (4 directions + zoom in/out + select) and a 15-action space (adding finer movement steps).
Reward Function: The reward is based on the positive improvement in the accumulated Intersection-over-Union (IoU) between the agent's selections and the ground-truth mask. A scaling parameter $S$ is introduced to modulate the reward, encouraging late-stage refinement when $S > 1$ .
Training Stability: To address optimization instabilities caused by batch normalization and dropout in RL, the authors employ MDR (a method for stable RL optimization under these regularizations).

Key Contributions

Interactive Environment Formulation: The work introduces a framework where the full WSI pyramid is treated as the interactive environment, enabling an agent to navigate via movement and zooming actions directly on the slide, rather than selecting from a pre-extracted set of patches.
End-to-End Sequential Segmentation: The authors train an end-to-end actor-critic agent to progressively construct a tumour segmentation mask through sequential interaction, eliminating the need for exhaustive patch extraction.
Comprehensive Analysis: The study provides extensive analysis of reward design (scaling parameter $S$ ), action-space granularity, architectural choices (backbone capacity, dropout), and training strategies (MDR, cohort extension).
Efficiency Comparison: The framework is benchmarked against patch-based classification pipelines (using the UNI foundation model) across both segmentation performance and inference efficiency.

Results
Experiments were conducted on pulmonary adenocarcinoma WSIs from a cohort of 94 patients.

Performance: The proposed RL agents achieved coarse segmentation quality comparable to patch-based approaches operating at similar magnification levels (e.g., Agent-7-Res18 achieved a Dice score of 0.80, comparable to UNI-4096). However, a performance gap remained in boundary agreement (NSD metrics), with agents achieving lower scores than patch-based ensembles (e.g., UNI-ENS). The authors note that errors were predominantly localized near tumour boundaries, a limitation attributed to the coarse spatial granularity of the action space and the limited depth of the pyramid explored.
Inference Efficiency: The RL framework demonstrated substantial improvements in inference speed. While patch-based methods required tens of seconds to minutes per slide (e.g., UNI-1024 averaged ~229s), the RL agents processed slides in approximately 3 seconds on average, with significantly lower variance.
Ablation Findings:
- Reward Scaling: Moderate values of $S$ (e.g., 4 or 16) provided a better trade-off between precision and recall compared to $S=1$ .
- Generalization: Extending the training cohort (using pseudo-labels for additional slides) significantly reduced the generalization gap, outperforming regularization techniques like dropout alone.
- Stability: MDR was essential for training architectures containing batch normalization and dropout, resolving severe instability observed with standard PPO.

Significance and Claims
The paper claims that modeling WSIs as interactive environments provides a promising direction for RL-based computational pathology. It demonstrates that direct sequential tumour segmentation on full slides is feasible under constrained interaction settings. The primary significance lies in achieving comparable coarse segmentation performance to patch-based baselines while drastically reducing inference time to a few seconds per slide. The authors position this work as a step toward clinically faithful, computationally efficient slide navigation that reflects the strategies of pathologists, though they acknowledge limitations regarding boundary precision and the need for explicit memory mechanisms in future work.

Technical Summary: Interactive Whole Slide Images for RL-based Tumour Segmentation

More like this