Related Experiment Video
Updated: Aug 23, 2025
09:24
Quantified Assessment of Infant's Gross Motor Abilities Using a Multisensor Wearable
Published on: May 17, 2024
1.5K
Semi-supervised body parsing and pose estimation for enhancing infant general movement assessment
Haomiao Ni1, Yuan Xue2, Liya Ma3
1College of Information Sciences and Technology, The Pennsylvania State University, University Park, PA, USA.
Medical Image Analysis
|November 3, 2022
Summary
This study introduces SiamParseNet (SPN), a semi-supervised model that improves early cerebral palsy (CP) detection using infant movement videos (IMVs). Augmenting video with body parsing and pose estimation significantly enhances general movement assessment (GMA) performance.
Area of Science:
- Computer Vision
- Machine Learning
- Developmental Pediatrics
Background:
- General movement assessment (GMA) of infant movement videos (IMVs) is crucial for early cerebral palsy (CP) detection.
- End-to-end neural networks show promise for GMA, but performance can be significantly improved by incorporating infant body parsing and pose estimation.
Purpose of the Study:
- To develop an efficient semi-supervised model (SiamParseNet - SPN) for utilizing partially labeled IMVs for infant body parsing.
- To enhance GMA performance by integrating body parsing and pose estimation with neural network-based video analysis.
- To explore data augmentation techniques for improving model training on limited labeled data.
Main Methods:
- Proposed SiamParseNet (SPN), a semi-supervised model with two branches for intra-frame segmentation and inter-frame label propagation, trained jointly on labeled and unlabeled IMVs.
- Introduced Factorized Video Generative Adversarial Network (FVGAN) for synthesizing novel labeled frames to augment training data.
- Employed a multi-source inference mechanism for robust testing, combining segmentation and propagation results.
Main Results:
- SPN coupled with FVGAN achieved state-of-the-art performance in infant body parsing on partially labeled IMVs.
- SPN demonstrated adaptability and superior performance for infant pose estimation.
- SPN models generalized well to a new clinical IMV dataset and significantly improved CRNN-based GMA prediction when combined with raw video inputs.
Conclusions:
- The proposed SiamParseNet (SPN) effectively utilizes partially labeled infant movement videos for body parsing and pose estimation.
- Integrating body parsing and pose estimation significantly enhances the performance of neural networks for general movement assessment (GMA) in early cerebral palsy (CP) detection.
- The developed methods show strong potential for clinical application in improving early diagnosis of neurological conditions in infants.
