Table 3. Perceptual consistency of histologic subtype in reconstructed images across cancer types.

Correlation between predictions of histologic subtype from real and reconstructed tiles, averaged per patient, across cancer types, demonstrating a high perceptual similarity of the histologic subtype of the real and generated images. For the TCGA datasets, a deep learning model was trained to predict subtype from real tiles for each cancer type using threefold cross-validation. The correlation between predictions for real/generated images is aggregated for the three held-out validation sets. For CPTAC and UCMC validation, a deep learning model trained across the entire corresponding TCGA dataset was used to generate predictions. AUROC and AP are listed for prediction of histologic subtype using the above models, as well as when predictions from these models are made with reconstructed (Gen) versions of tiles. The similar AUROC/AP from real tiles and reconstructed tiles illustrates the reconstructed tiles retain informative data with regard to histologic subtype.

Source		n	Histology (%)	Pearson r	P value, correlation	AUROC	AUROC, Gen.	AP	AP, Gen.
TCGA	BRCA	734	Ductal (77.2) Lobular (22.8)	0.94 (0.93–0.95)	<1 × 10−99	0.96	0.96	0.92	0.90
TCGA	LUNG	941	Adeno (49.6) Squamous (51.4)	0.94 (0.92–0.96)	9.57 × 10−72	0.97	0.98	0.94	0.95
TCGA	ESCA	147	Adeno (44.9) Squamous (55.1)	0.95 (0.93–0.96)	<1 × 10−99	0.99	0.99	0.96	0.94
TCGA	KIDNEY	363	Clear (74.4) Papillary (25.6)	0.91 (0.9–0.91)	<1 × 10−99	0.95	0.95	0.91	0.89
CPTAC	BRCA	92	Ductal (92.4) Lobular (7.6)	0.64 (0.5–0.75)	5.80 × 10−12	0.69	0.39	0.57	0.13
CPTAC	LUNG	415	Adeno (49.3) Squamous (50.7)	0.94 (0.93–0.95)	<1 × 10−99	0.96	0.96	0.92	0.90
UCMC	BRCA	820	Ductal (85.5) Lobular (14.5)	0.94 (0.93–0.95)	<1 × 10−99	0.88	0.79	0.59	0.46