Edit this page

ProjectWeek

Back to Projects List

Integration of Diabetic Foot Segmentation Algorithm based on Deep Learning

Key Investigators

Project Description

This project is the follow-up of the development of a module to use InfraRed (IR) sensors in 3D Slicer for medical diagnosis, intended to use for monitoring of foot ulcers in diabetic patients, presented on the 28thPW NA-MIC. This project is proposed as a new stage in the diabetic foot assessment previously worked.

The aim is to integrate an algorithm, which is based on Deep Learning, for foot segmentation using multimodal images (visible and depth-map images). The resulting mask will be applied on thermal images in order to analyze the temperature pattern and detect possible foot ulcers.

Objective

  1. Update the “Diabetic Foot” extension created on the 28thPW NA-MIC.
  2. Integrate the foot segmentation algorithm presented on the paper [1].

Approach and Plan

  1. Integrate TorchScript models, an intermediate representation of a PyTorch model
  2. Include a point cloud processing library
  3. Implementing the foot segmentation algorithm
  4. (Optional) Visualization of point cloud using VTK

Progress and Next Steps

Illustrations

Workflow Proposed workflow for feet segmentation where the images are represented as squares and the operations as rectangles.The network prediction is used to set the ROI on the depth image, the point cloud, and then a second segmentation was applied to extract geometry models, specifically planes.

Progress

Premasked Original depth map. This point cloud is dense populated and contains several noisy points.




ROI Region of interest obtained from the segmentation of Deep Learning in the RGB image. The number of points has been reduced, but still contains noise points.




Compare Comparison between the two images mentioned above.




ResultPCD Feet Segmentation module result. The point cloud visualization is just representative and that result has to be represented as image (Work in Progress) to be used as a mask in the thermal image.

Background and References

[1] Hernández, A., Arteaga-Marrero, N., Villa, E., Fabelo, H., Callicó, G. M., & Ruiz-Alzola, J. (2019, September). Automatic Segmentation Based on Deep Learning Techniques for Diabetic Foot Monitoring Through Multimodal Images. In International Conference on Image Analysis and Processing (pp. 414-424). Springer, Cham. Avalible from: https://link.springer.com/chapter/10.1007%2F978-3-030-30645-8_38

[2] TorchScript tutorial

[3] Diabetic Foot Extension. Repository with the extension that includes the “SuperBuild” option and the developed module.