Spatial-temporal Hierarchical Reinforcement Learning for Interpretable Pathology Image Super-Resolution

Wenting Chen, Jie Liu, Tommy W. S. Chow, Yixuan Yuan

Introduction

High-resolution histopathology slides are the cornerstone of modern cancer diagnosis, providing the cellular detail necessary for identifying malignancy. However, the acquisition of such images is often bottlenecked by the high cost and slow throughput of high-end scanning equipment. Digital Super-Resolution (SR) offers a potential software-based solution to "up-sample" lower-quality scans. Yet, traditional deep learning approaches, particularly Generative Adversarial Networks (GANs), suffer from a "black-box" nature. They are prone to generating "hallucinations"—plausible-looking but biologically inaccurate artifacts—that can mislead clinical judgment. Furthermore, these models are computationally expensive, treating simple background pixels and complex nuclear structures with the same level of intensity.

To address these challenges, the STAR-RL (Spatial-Temporal hierARchical Reinforcement Learning) framework is proposed. STAR-RL reformulates image restoration as a transparent, adaptive process that prioritizes diagnostic relevance and interpretability.

Methodology: The STAR-RL Framework

The core innovation of STAR-RL lies in its treatment of super-resolution as a Markov Decision Process (MDP). Rather than a single-pass inference through a deep neural network, the model employs a hierarchical structure of three specialized agents to restore an image iteratively:

  • Spatial Manager (spM): This high-level agent identifies regions of interest within a whole slide image. It learns to focus on "corrupted" or biologically dense patches (such as those containing cell nuclei) while bypassing simpler regions like empty background or uniform cytoplasm.
  • Temporal Manager (tpM): This agent monitors the restoration progress of a specific patch. By determining when a patch has reached a sufficient level of quality, it triggers an "early exit," significantly reducing redundant computations and preventing over-processing artifacts.
  • Patch Worker (PW): The Patch Worker operates at the pixel level. Crucially, instead of generating pixel values directly, it selects actions from a library of ten predefined, classical image filters (e.g., Sobel, Laplace, Gaussian).

This hierarchical approach ensures that the final high-resolution output is the result of a sequence of understandable mathematical operations. This transparency allows pathologists to verify the restoration process, mitigating the "black-box" concerns associated with standard SR models.

Data and Evaluation

The framework was rigorously tested across three major pathology benchmarks:

  • HistoSR: A large-scale dataset of 35,000 images used for primary training and benchmarking against state-of-the-art SR methods.
  • PCam (PatchCamelyon): Utilized to evaluate the downstream impact of SR on automated tumor recognition tasks.
  • Camelyon16: Used to demonstrate the scalability of STAR-RL on gigapixel-scale Whole Slide Images (WSIs).

Results and Discussion

STAR-RL demonstrates that interpretability does not have to come at the cost of performance. On the HistoSR benchmark, the model achieved a Peak Signal-to-Noise Ratio (PSNR) of 31.403, surpassing established models like PathSRGAN (31.196) and SWD-Net (31.368).

Beyond pixel-level metrics, the model's clinical utility was validated through tumor classification tasks. Using STAR-RL as a pre-processing step improved tumor recognition accuracy to 96.41%, compared to 93.06% using original low-resolution images. This suggests that the model effectively restores the structural details most relevant to diagnosis.

Furthermore, STAR-RL is exceptionally efficient. With a parameter count of only 0.5 MB, it is significantly more lightweight than its competitors. Its hierarchical nature allows it to allocate computational resources dynamically, focusing on complex biological structures while maintaining high visual fidelity and reducing reconstruction errors in dense cellular regions.

Conclusion and Future Work

STAR-RL represents a significant step toward trustworthy AI in digital pathology. By combining the strengths of reinforcement learning with the transparency of classical image processing, it provides a high-performance, interpretable, and efficient solution for image super-resolution.

Future research will focus on expanding the model's scope from local patch enhancement to global slide-level processing. Additionally, exploring self-supervised learning techniques could remove the dependency on paired high-resolution/low-resolution training data, allowing the model to learn directly from unlabelled clinical archives.

Summary updated

Abstract

Pathology image are essential for accurately interpreting lesion cells in cytopathology screening, but acquiring high-resolution digital slides requires specialized equipment and long scanning times. Though super-resolution (SR) techniques can alleviate this problem, existing deep learning models recover pathology image in a black-box manner, which can lead to untruthful biological details and misdiagnosis. Additionally, current methods allocate the same computational resources to recover each pixel of pathology image, leading to the sub-optimal recovery issue due to the large variation of pathology image. In this paper, we propose the first hierarchical reinforcement learning framework named Spatial-Temporal hierARchical Reinforcement Learning (STAR-RL), mainly for addressing the aforementioned issues in pathology image super-resolution problem. We reformulate the SR problem as a Markov decision process of interpretable operations and adopt the hierarchical recovery mechanism in patch level, to avoid sub-optimal recovery. Specifically, the higher-level spatial manager is proposed to pick out the most corrupted patch for the lower-level patch worker. Moreover, the higher-level temporal manager is advanced to evaluate the selected patch and determine whether the optimization should be stopped earlier, thereby avoiding the over-processed problem. Under the guidance of spatial-temporal managers, the lower-level patch worker processes the selected patch with pixel-wise interpretable actions at each time step. Experimental results on medical images degraded by different kernels show the effectiveness of STAR-RL. Furthermore, STAR-RL validates the promotion in tumor diagnosis with a large margin and shows generalizability under various degradations. The source code is available at https://github.com/CUHK-AIM-Group/STAR-RL.
BibTeX
@article{chen2024spatial,
  author = {Wenting Chen and Jie Liu and Tommy W. S. Chow and Yixuan Yuan},
  title = {{Spatial-temporal Hierarchical Reinforcement Learning for Interpretable Pathology Image Super-Resolution}},
  journal = {IEEE Transactions on Medical Imaging},
  year = {2024},
  url = {https://openreview.net/forum?id=biD6JLd27t},
}
Figure