MuscleMoDeL
This project aims at developing an artificial intelligence-based approach for the automated extraction of patient-specific information on spinal muscle morphology from medical images.
Factsheet
- Lead school School of Health Professions
- Additional schools School of Engineering and Computer Science
- Institute Physiotherapy
- Research unit Spinal Movement Biomechanics
- Funding organisation BFH
- Duration 01.01.2021 - 31.07.2022
- Project management Dr. Stefan Schmid
- Head of project Dr. Stefan Schmid
-
Project staff
Prof. Marcus Hudritsch
Benjamin Noah Fankhauser
Christian Bangerter
Michael von Arx - Partner Inselspital
- Keywords Medical image processing, Neural networks, U-Net, CT scans
Situation
MSK modeling is only accurate if the models are patient-specific, particularly in terms of muscle properties. To adjust such parameters, we usually rely on manually segmented CT scans, but these processes are highly time-consuming and therefore not applicable for research on larger samples or for everyday clinical practice. A possible way to speed up these processes might be to use an artificial intelligence (AI) approach such as deep learning, which uses biologically-inspired neural networks that enable computers to learn from observational data. A well trained neural network could thereby reduce the time needed for image segmentation from several hours to a couple of seconds per patient. The overall aim of the project is to create, train and verify a neural network for the automated ex-traction of patient-specific paraspinal muscle morphology from CT scans to allow for appropriate simulation of spinal loading using MSK modeling.
Course of action
To be able to train the neural network appropriately, we will establish a training dataset consisting of approximately 200 trunk CT/MRI scans of healthy children and adolescents. To extract the required training data, scans will be manually segmented, including the assessment of trunk muscle morphology by determining muscle cross-sectional area (CSA) and relative location for each horizontal vertebral mid-plane using a custom-built image processing software. Using our training dataset, we will then retrain a ready trained convolutional neural network (CNN) for the recognition of vertebral body as well as erector spinae and multifidi muscle contours. The process of retraining a ready trained network is called transfer learning and takes far fewer epochs and results in better accuracy. Finally, muscle morphology data will be implemented into our MSK models using a previously established MATLAB-pipeline.
Result
A dataset of 56 (instead of the planned 200) CT scans was established. To be able to appropriately segment the CT scans, we first developed a custom segmentation tool, which we used to manually segment the scans (952 individual slices). We then used the data from 54 selected scans to train an extended U-Net architecture with 3 layers, first for detecting vertebral bodies and then for muscle cross-sections. Finally, the network was evaluated using data from the 2 remaining scans, which resulted in a dice score of 0.84 and an intersection over union of 0.62. Hence, the network was able to automatically detect the selected muscle cross-sections, but not yet with human-like performance.
Looking ahead
This project laid the groundwork for a fully automated recognition of muscle cross-sections to inform MSK models of the spine. To achieve human-like performance of the neural network, however, further optimizations are required.
