1 of 45

​

​

​

Making AI in medical imaging

accessible and reproducible

github.com/MHubAI

mhub.ai

2 of 45

The Problem | Framework & Dependencies

3 of 45

The Problem | Framework & Dependencies

4 of 45

The Problem | Framework & Dependencies

​

5 of 45

The Problem | Framework & Dependencies

6 of 45

The Problem | Input Data, Format & Structure

CT Scan Image

7 of 45

The Problem | Input Data, Format & Structure

8 of 45

The Problem | Input Data, Format & Structure

9 of 45

The Problem | Input Data, Format & Structure

10 of 45

The Problem | Alignment of Output Data

11 of 45

The Problem | Alignment of Output Data

12 of 45

The Problem | Alignment of Output Data

13 of 45

The Problem | Alignment of Output Data

LEFT_KIDNEY

left kidney

14 of 45

The Problem | Summary

So what do we need to �Run such a pipeline

  • Model Information
  • Environment Setup
  • Input Format
  • Input Structure
  • Output Format
  • Output Structure

15 of 45

The Problem | Summary

So what do we need to �Run such a pipeline

  • Model Information
  • Environment Setup
  • Input Format
  • Input Structure
  • Output Format
  • Output Structure

Can we find a solution to simplify the setup of DL models in medical imaging and make the process scalable?

16 of 45

Our Solution | MHub

17 of 45

Our Solution | MHub

Licence must at least allow for educational use.

18 of 45

Our Solution | MHub

​

19 of 45

Our Solution | MHub

​

All code for Model A is cloned from their public repository

20 of 45

Our Solution | MHub

21 of 45

Our Solution | MHub

The weights �for Model A are downloaded from a version controlled resource.

22 of 45

Our Solution | MHub

MHub-IO Framework provides a toolbox of sequential executable modules, e.g.

  • Import
  • Convert
  • Filter
  • Report
  • Organize

23 of 45

Our Solution | MHub

MHub-IO Framework provides a toolbox of sequential executable modules, e.g.

​

The AI Pipeline �(Model A) is wrapped as-is in such an MHub-IO Module.

24 of 45

Our Solution | MHub

25 of 45

Our Solution | MHub

26 of 45

Our Solution | MHub

27 of 45

Our Solution | MHub

28 of 45

Our Solution | MHub

29 of 45

Our Solution | MHub

30 of 45

Our Solution | MHub Default Workflow

Every MHub model has a default workflow starting from DICOM.

→ standardized input

​

All segmentation models produce DICOMSEG as an output.

→ standardized output

31 of 45

MHub - Model Repository

32 of 45

MHub - Model Repository

  • Dockerized containers require no environment setup and are platform agnostic.�
  • Search and explore models on our website with search and filter functionality.�
  • Model serving, standardization, and the I/O framework are all significant improvements from the old platform!

33 of 45

MHub - Model Repository

​

​

​

​

34 of 45

MHub - 1 Line to Run a Model

35 of 45

MHub - Model Repository

​

​

​

​

36 of 45

MHub - Model Repository

​

​

​

​

37 of 45

MHub - Model Repository

​

​

​

​

38 of 45

MHub - Model Repository

​

​

​

​

39 of 45

MHub - Model Repository

​

​

​

​

mhubai/totalsegmentator:latest

40 of 45

MHub - Model Repository

​

​

​

​

mhubai/platipy:latest

41 of 45

MHub - Model Repository

​

​

​

​

mhubai/lungmask:latest

42 of 45

MHub - Model Repository

​

​

​

​

mhubai/gc_lunglobes:latest

43 of 45

MHub - Made for Integration

44 of 45

MHub - Made for Integration

​

​

​

​

  • Containerization �→ platform agnostic �
  • Standardized I/O�→ Adapt to the MHub standard API�
  • Standardized Config�→ Customize MHub workflows

45 of 45

Acknowledgements

​

​

​

We acknowledge the financial support received for this work from the National Institute of Health (NIH) and the European Research Council (ERC).

​

​

​

​

​

​

Special thanks are extended to Dr. Andrey Fedorov and Dr. Hugo Aerts for their significant engagement and invaluable contributions to the project. Their expertise and commitment have been instrumental in advancing MHub to its current stage. ��We would also like to express our sincere gratitude to Dr. Keyvan Farahani, Dr. Linmin Pei, and Dr. Ulrike Wagner for their steadfast support and assistance throughout our research.