RSOM epidermis and vessel segmentation in pytorch

Preface

This reposity contains the python source code to perform segmentation of raster-scan optoacoustic mesoscopy (RSOM) data (Aguirre J. et al., Precision assessment of label-free psoriasis biomarkers with ultra-broadband optoacoustic mesoscopy, Nat Biomed Eng 1, 0068 (2017) )

Setup

This repository was tested on Ubuntu 16.04.6 LTS and 18.04.3 LTS.

It uses git large file storage. In order to work it correctly you need to install git lfs following this manual for git lfs. As an alternative, you can clone the repository as usual, and download the large files in ./data/input and ./data/models manually.

A GPU with 16GiB memory is highly recommended.

The main python dependencies are

• nibabel
• torch

However, for a complete list, see and run

pip install -r requirements.txt

Usage

In ./data/input there are example matlab files, which serve as input. Files must follow a naming scheme similar to R_<datetime>_<letter-number identifier>_{LF,HF}.mat. Surface files, if present, must be named Surf_<datetime>.mat. Currently there is no support for other input file types, however, writing a wrapper for any needed file format can be done with little effort.

To run the epidermis and vessel segmentation, run

$ python pipeline.py

If you want to change input and output files, or use cpu for computation, edit pipeline.py accordingly.

If you have a smaller GPU, you can edit variable divs=(1,1,2), to divs=(1,2,2), or even larger values. divs are the numbers, the input tensor is split in each dimension, and closely related to the amount of memory required.

Results

After successful segmentation, results are placed in ./data/output/. Furthermore, all tensors of all segmentation steps are placed in ./data/output/tmp. If one wishes to remove intermediate tensors, you can adjust that in pipeline.py.

Maximum intensity projection of

• input RSOM data
• RSOM data and segmentation result of vessels (blue) and epidermis (white).

References

[1] Ronneberger, O., Fischer, P., Brox, T., U-Net: Convolutional networks for biomedical image segmentation.

[2] Tetteh G. et al., DeepVesselNet: Vessel Segmentation, Centerline Prediction, and Bifurcation Detection in 3-D Angiographic Volumes.

[3] Aguirre J. et al. Precision assessment of label-free psoriasis biomarkers with ultra-broadband optoacoustic mesoscopy, Nat Biomed Eng 1, 0068 (2017)