Multicenter Computer-Aided Diagnosis for Lymph Nodes Using Unsupervised Domain-Adaptation Networks Based on Cross-Domain Confounding Representations

Figure 2

Illustration of the proposed network architecture. (a) Domain confounding representation through cross-domain mapping: the encoder F and the decoder G constitute the VAE architecture for unsupervised representation learning. The D module constitutes the GAN discriminator, while the C module constitutes a classifier. The encoder F uniformly encodes images from two domains. Paired decoders process different domain features, enabling cross-domain pixel-level image reconstruction and adversarial discrimination. (b) Classification cycle consistency: the reconstructed image based on source-domain features, as shown by the black line, will be constrained by classification cycle consistency through F and C. (c) Illustration of the loss overview.

(a)

(b)

(c)