deepMTJ
is a machine learning approach for automatically tracking of muscle-tendon junctions (MTJ) in ultrasound images. Our method is based on a convolutional neural network trained to infer MTJ positions across various ultrasound systems from different vendors, collected in independent laboratories from diverse observers, on distinct muscles and movements. We built deepMTJ
to support clinical biomechanists and locomotion researchers with an open-source tool for gait analyses.
This repository contains the full test dataset used for deepMTJ
performance assessments, the trained TensorFlow (Keras) model and a Link to the code repository of deepMTJ. Furthermore, we provide online predictions using deepMTJ
via a (For multiple and large file predictions) and via deepmtj.org (Cloud based predictions).
The dataset comprises 1344 images of muscle-tendon junctions recorded with 3 ultrasound imaging systems (Aixplorer V6, Esaote MyLab60, Telemed ArtUs), on 2 muscles (Lateral Gastrocnemius, Medial Gastrocnemius), and 2 movements (isometric maximum voluntary contractions, passive torque movements).
We have included the ground truth labels for each image. These reference labels are the computed mean from 4 specialist labels. Specialist annotators had 2-10 years of experience in biomechanical and clinical research investigating muscles and tendons in 2-9 ultrasound studies in the past 2 years.
Paper | Code | Results | Date | Stars |
---|