Towards a Privacy-Preserving Distributed Cloud Service for Preprocessing Very Large Medical Images

Yuandou Wang, Neel Kanwal, Kjersti Engan, Chunming Rong, Zhiming Zhao

Abstract

Digitized histopathology glass slides, known as Whole Slide Images (WSIs), are often several gigapixels large and contain sensitive metadata information, which makes distributed processing unfeasible. Moreover, artifacts in WSIs may result in unreliable predictions when directly applied by Deep Learning (DL) algorithms. Therefore, preprocessing WSIs is beneficial, e.g., eliminating privacy-sensitive information, splitting a gigapixel medical image into tiles, and removing the diagnostically irrelevant areas. This work proposes a cloud service to parallelize the preprocessing pipeline for large medical images. The data and model parallelization will not only boost the end-to-end processing efficiency for histological tasks but also secure the reconstruction of WSI by randomly distributing tiles across processing nodes. Furthermore, the initial steps of the pipeline will be integrated into the Jupyter-based Virtual Research Environment (VRE) to enable image owners to configure and automate the execution process based on resource allocation.

The Challenge of Gigapixel Pathology

In digital pathology, Whole Slide Images (WSIs) are the gold standard for diagnosis and research. However, their sheer scale—often reaching several gigapixels—presents a significant computational hurdle. Before these images can be utilized by AI-driven diagnostic tools, they must undergo extensive preprocessing to remove artifacts such as air bubbles, blurred regions, or tissue folds. These artifacts, if left unaddressed, can lead to "garbage in, garbage out" scenarios in downstream analysis.

The artifact detection pipeline: identifying and segmenting diagnostically irrelevant areas like blood, blur, and air bubbles in a WSI.

Beyond the technical challenge of processing speed, two major barriers exist:

  1. Infrastructure Constraints: Most clinical settings lack the massive GPU clusters required to process thousands of giant images efficiently.
  2. Privacy Concerns: Uploading sensitive medical data to public clouds is risky. WSIs contain metadata and spatial structures that could potentially identify patients, making standard cloud processing a security liability.

A Distributed, Privacy-Preserving Methodology

To address these issues, this work proposes a five-step distributed framework integrated into a Jupyter-based Virtual Research Environment (VRE). This system shuffles image tiles across distributed cloud nodes, ensuring that no single processing unit ever sees the complete, identifiable image.

The proposed five-step workflow for distributed WSI preprocessing, including data preparation, resource allocation, and secure aggregation.

The pipeline consists of the following stages:

  1. Data & Workflow Preparation: Metadata is stripped for anonymity. The WSI is split into tiles, and their spatial relationships are encoded in a coordinate matrix ($A_x$) to prevent unauthorized reconstruction.
  2. Resource Allocation: The system maps tasks to cloud resources using bi-objective optimization, balancing execution time (makespan) against monetary costs.
  3. Deployment & Execution: Containerized microservices process the tiles. By distributing tiles randomly across nodes, the full image remains protected even if a node is compromised.
  4. Load Balancing: Task-based load balancing ensures high throughput and prevents slow nodes from delaying the overall pipeline.
  5. Data Aggregation: Artifact detection results from an ensemble of five Deep Learning models are sent back to a trusted server. The final segmentation mask is reconstructed using random-value perturbation for added security.

Key Innovations

  • Security through Fragmentation: Privacy is achieved by "shredding" the image. An attacker breaching a single node only accesses disconnected, anonymous tissue squares rather than a complete patient record.
  • Microservice Architecture: Artifact detection is treated as a set of fine-grained, reusable services (e.g., specific models for blood or blur), allowing researchers to swap models or scale specific components as needed.
  • Bi-Objective Optimization: A mathematical model solves the "Cloud Dilemma," calculating the optimal balance between processing speed and research budget by evaluating the unit price of infrastructure against task makespan.

Conclusion and Future Work

This methodology provides a blueprint for a privacy-preserving computational pathology pipeline. By integrating these tools into a Virtual Research Environment (VRE), the framework allows pathologists to manage complex cloud executions without deep expertise in distributed systems.

As a work-in-progress, future efforts will focus on real-world validation across heterogeneous cloud environments and developing a full prototype to quantify the speedup and cost-savings compared to traditional single-machine preprocessing. The team also aims to validate the pipeline's cross-institutional utility using datasets sourced from multiple different hospitals.

Summary updated

Abstract

Digitized histopathology glass slides, known as Whole Slide Images (WSIs), are often several gigapixels large and contain sensitive metadata information, which makes distributed processing unfeasible. Moreover, artifacts in WSIs may result in unreliable predictions when directly applied by Deep Learning (DL) algorithms. Therefore, preprocessing WSIs is beneficial, e.g., eliminating privacy-sensitive information, splitting a gigapixel medical image into tiles, and removing the diagnostically irrelevant areas. This work proposes a cloud service to parallelize the preprocessing pipeline for large medical images. The data and model parallelization will not only boost the end-to-end processing efficiency for histological tasks but also secure the reconstruction of WSI by randomly distributing tiles across processing nodes. Furthermore, the initial steps of the pipeline will be integrated into the Jupyter-based Virtual Research Environment (VRE) to enable image owners to configure and automate the execution process based on resource allocation.
BibTeX
@misc{wang2023towards,
  author = {Yuandou Wang and Neel Kanwal and Kjersti Engan and Chunming Rong and Zhiming Zhao},
  title = {{Towards a Privacy-Preserving Distributed Cloud Service for Preprocessing Very Large Medical Images}},
  year = {2023},
  eprint = {2307.06266},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2307.06266},
}
Figure