metadata

language:
  - en
pipeline_tag: image-segmentation
tags:
  - medical
library_name: fastMONAI

Multi-Center CNN-based spine segmentation from T2w MRI using small amounts of data

Pre-trained weights and exported models for our spine segmentation project. The source code, designed to reproduce our test results and facilitate training and running inference on your own data, is available on GitHub: https://github.com/MMIV-ML/fastMONAI/tree/master/research

Requirements

Last checked and validated with fastMONAI version 0.3.2. Please ensure that you have the correct version of fastMONAI installed to guarantee the correct operation of the model.

Data

Haukeland University Hospital (HUH): 15 T2-weighted MRI scans

Study

Scanner Specifications

Type: Siemens MAGNETOM Avanto 1.5 T
Slice Thickness: 4 mm
Field of View (FOV): 300x300 mm^2
Matrix: 384x269
Interslice Gap: 0.4 mm
Repetition Time: 3700 ms
Echo Time: 87 ms
Number of Acquisitions: 2

Sahlgren-ska University Hospital (SUH): 10 T2-weighted MRI scans

Study

Scanner Specifications

Type: Siemens MAGNETOM Aera 1.5 T
Slice Thickness: 3.5 mm
Field of View (FOV): 300x300 mm^2
Matrix: 384x384
Interslice Gap: 0.7 mm
Repetition Time: 3500 ms
Echo Time: 95 ms
Number of Acquisitions: 1

Data Annotation

Process: Manual segmentation of vertebrae
Experts: Conducted by an MRI expert with consensus from a senior radiologist
Software: ITK-SNAP
Scope: Segmented vertebrae in all slices, including bone marrow but excluding vertebral cortex

External Data Evaluation

Source: Annotated T2-weighted MR images of the Lower Spine
Scan Specifications: 23 scans, 1.5 T in sagittal orientation
Segments: Seven vertebrae are manually segmented on each image (Th11–L5)
Modification: Th11 manually removed for consistency with the annotation for SUH and HUH.