Constellation loss is a function that optimizes a deep learning classifier with very few training images. It goes one step further than other loss functions by simultaneously learning distances among all class combinations. It can attract same class image embeddings whilst pulling apart the rest of the classes, all at the same time. This way, an optimal embedding or descriptor of the image is generated.

This approach can be used on different domains where data is scarce and target labels not available, for example when adopting novel imaging techniques as OCT or MPT. Also facilitates the implementation of image retrieval applications based on clinical features thanks to metric learning. Besides, it offers a more efficient and cheaper training process.

The details of this output are openly accessible through publications in scientific journals at the scientific community´s disposal

CONTACT

Artzai Picón

Tecnalia Research & Innovation (TECNALIA)

www.tecnalia.com

artzai.picon@tecnalia.com