Fig. 2: Similarity measure between feature maps of real modalities (obtained in training Stage I), i.e., PETFDG and PETAV45, and the corresponding synthesized ones (obtained in training Stage II) on test data.
a Boxplots of cosine similarity (CS) and KL divergence (KLD) of the 16-channel backbone features between real and synthesized modalities for normal cognition (NC) vs. Alzheimer’s disease (AD) classification across the ADNI test data, with the median values showing in orange bar. The barplots show the average CS of the 16 channels of the classification backbone (average value of each channel as shown in boxplot (a) in blue) and additional two bars (in dark green) for the average CS of the features of the two fully connected layers in the classification head. b T-SNE visualization of feature representations of real and synthesized modalities for NC vs. AD, static mild cognitive impairment (sMCI) vs. progressive MCI (pMCI), subcortical vascular disease with no cognitive impairment (NCI) vs. subcortical vascular mild cognitive impairment (svMCI), and MGMT promoter methylation status (methylated or unmethylated) classifications by t-SNE plots. Features are collected from the output of the classification backbone. c T-SNE visualization of feature representations of the acquired T1w and synthesized other modalities. It turns out that although synthesized features are generated by T1w, they are complementary to T1w (containing modality-specific patterns).
