Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Stroke Lesion Segmentation in Clinical Workflows: A Modular, Lightweight, and Deployment-Ready Tool

Yann Kerverdo, Florent Leray, Youwan Mahé, Stéphanie Leplaideur, Francesca Galassi

TL;DR

StrokeSeg addresses the challenge of deploying high-performing stroke lesion segmentation models in clinical workflows by rearchitecting nnU-Net pipelines into a modular, lightweight tool. It leverages ONNX Runtime with Float16 quantisation, decouples preprocessing, inference, and postprocessing, and packages the solution for Windows and Linux. The approach preserves accuracy (mean Dice difference $<10^{-3}$) while substantially reducing package size (about $60 ext{%}$) and enabling rapid GPU-accelerated inference. This work demonstrates that clinically usable, portable segmentation tools can be built from research-grade models, with implications for broader adoption of automated stroke lesion quantification in clinical practice.

Abstract

Deep learning frameworks such as nnU-Net achieve state-of-the-art performance in brain lesion segmentation but remain difficult to deploy clinically due to heavy dependencies and monolithic design. We introduce \textit{StrokeSeg}, a modular and lightweight framework that translates research-grade stroke lesion segmentation models into deployable applications. Preprocessing, inference, and postprocessing are decoupled: preprocessing relies on the Anima toolbox with BIDS-compliant outputs, and inference uses ONNX Runtime with \texttt{Float16} quantisation, reducing model size by about 50\%. \textit{StrokeSeg} provides both graphical and command-line interfaces and is distributed as Python scripts and as a standalone Windows executable. On a held-out set of 300 sub-acute and chronic stroke subjects, segmentation performance was equivalent to the original PyTorch pipeline (Dice difference $<10^{-3}$), demonstrating that high-performing research pipelines can be transformed into portable, clinically usable tools.

Stroke Lesion Segmentation in Clinical Workflows: A Modular, Lightweight, and Deployment-Ready Tool

TL;DR

StrokeSeg addresses the challenge of deploying high-performing stroke lesion segmentation models in clinical workflows by rearchitecting nnU-Net pipelines into a modular, lightweight tool. It leverages ONNX Runtime with Float16 quantisation, decouples preprocessing, inference, and postprocessing, and packages the solution for Windows and Linux. The approach preserves accuracy (mean Dice difference ) while substantially reducing package size (about ) and enabling rapid GPU-accelerated inference. This work demonstrates that clinically usable, portable segmentation tools can be built from research-grade models, with implications for broader adoption of automated stroke lesion quantification in clinical practice.

Abstract

Deep learning frameworks such as nnU-Net achieve state-of-the-art performance in brain lesion segmentation but remain difficult to deploy clinically due to heavy dependencies and monolithic design. We introduce \textit{StrokeSeg}, a modular and lightweight framework that translates research-grade stroke lesion segmentation models into deployable applications. Preprocessing, inference, and postprocessing are decoupled: preprocessing relies on the Anima toolbox with BIDS-compliant outputs, and inference uses ONNX Runtime with \texttt{Float16} quantisation, reducing model size by about 50\%. \textit{StrokeSeg} provides both graphical and command-line interfaces and is distributed as Python scripts and as a standalone Windows executable. On a held-out set of 300 sub-acute and chronic stroke subjects, segmentation performance was equivalent to the original PyTorch pipeline (Dice difference ), demonstrating that high-performing research pipelines can be transformed into portable, clinically usable tools.

Paper Structure

This paper contains 14 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Architecture overview
  4. Scientific package
  5. Performance-oriented inference
  6. Packaging and distribution
  7. Main packaging target
  8. Secondary packaging targets.
  9. Software design and extensibility
  10. Results
  11. Accuracy
  12. Performance and size
  13. Discussion and Conclusion
  14. Compliance with ethical standards

Figures (2)

  • Figure 1: Challenges at the intersection of research, engineering, and clinical practice. Each overlap highlights domain-specific constraints addressed in this work.
  • Figure 2: Overview of the StrokeSeg packaging workflow. The CMake project compiles the C++ launcher and gathers the scientific Python modules, embedded Python environment, and pretrained models. CPack then builds a platform-specific installer (e.g., NSIS for Windows) supporting both graphical and command-line execution. The resulting package can be launched directly by end users without manual environment configuration.