Sentence embeddings for missing clinical metadata
The Metadata Processing Block (MetaBlock) is a strong multimodal fusion mechanism, but it consumes metadata as independent categorical features. Nothing in that representation knows that two attributes are semantically related, and a missing field is just a hole in the vector.
In deployment — particularly in the remote regions where CAD systems have the most value — incomplete records are the norm rather than the exception.
A sentence embedding algorithm turns the clinical attributes into dense vectors that encode the semantic relationships between features, rather than treating each one as an isolated dimension.
The change is confined to how metadata enters the fusion block, so the method inherits MetaBlock's architecture and can be swapped into an existing pipeline.
Alongside PAD-UFES-20, the method is tested on a new extended version of the dataset roughly seven times its size, which is where robustness claims actually get tested.
Inference time and model size were treated as first-class results — the method is meant to sit inside a real CAD system, not only in a benchmark table.
ResNet-50 backbone, balanced accuracy.
Later reused as the fusion backbone for the LLM-generated verbose metadata study.
Bouzon, P. H. G., Rocha, W. F. da, Souza, L. A., Pacheco, A. G. C. MetaBlock-SE: A Method to Deal With Missing Metadata in Multimodal Skin Cancer Classification. IEEE Journal of Biomedical and Health Informatics, vol. 29, pp. 8855–8862, 2025.
The implementation is not public yet. If you would like to reproduce the experiments or discuss the setup, get in touch and I will share what I can.
Other parts of the same research line.