Fig. 5. Class and layer blending provides nuanced insights into BRAF^V600E-like and RAS-like thyroid neoplasms.

a cGANs can create synthetic layer-blended images by conditioning the network using different embeddings at each layer. b Layer blending with a seed from the Thyroid cGAN reveals different morphologic changes associated with the RAS-like and BRAF^V600E-like gene expression spectrum. Each image includes a corresponding classifier prediction, from -1 (BRAF^V600E-like) to +1 (RAS-like). Image B1 is a fully RAS-like image, and Image B6 is a fully BRAF^V600E-like image. Images B2-B5 are generated by using different class embeddings at each cGAN layer. Examining the resulting morphologic changes that occur when passing the BRAF^V600E-like embedding to different layers illustrates different types of morphologic changes associated with the BRAF^V600E-RAS spectrum.