Fig. 2: The overall architecture of MAM. | Nature Communications

Fig. 2: The overall architecture of MAM.

From: Automating General Movements Assessment with quantitative deep learning to facilitate early screening of cerebral palsy

Fig. 2

MAM consists of a Ref Branch, a Main Branch, and an Info Branch. The Ref Branch, containing a 3D pose estimation step, an input construction step, a spatio-temporal Transformer (Supplementary Fig. 1) and a classifier, employs small FMs or non-FMs clips as inputs, and outputs FMs and non-FMs probability. The Main Branch, with an additional split step and an attention-based fusion (Supplementary Fig. 2), employs whole videos as inputs, and outputs normal and risk probability. The Info Branch employs basic characteristics as input, and output normal and risk probability by a fully connected neural network. The normal and risk probability given by the Main Branch and Info Branch are combined to give the final normal and risk probability.

Back to article page