Table 2. Perceptual consistency of tumor grade in reconstructed images across cancer types.

Correlation between predictions of grade from real and reconstructed tiles, averaged per patient, across cancer types, demonstrating a high perceptual similarity of the grade of the real and generated images. For the TCGA datasets, a deep learning model was trained to predict grade 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 the CPTAC validation, a deep learning model trained across the entire corresponding TCGA dataset was used to generate predictions. Average area under the receiver operating characteristic (AUROC) and average precision (AP) are listed for prediction of grade 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 grade.

Source		n	High grade (%)	Pearson r	P value, correlation	AUROC	AUROC, Gen.	AP	AP, Gen.
TCGA	BRCA	943	36.5	0.89 (0.88–0.91)	< 1 × 10−99	0.81	0.74	0.69	0.60
TCGA	PAAD	168	29.8	0.85 (0.81–0.89)	3.90 × 10−49	0.52	0.58	0.32	0.40
TCGA	PRAD	227	22.9	0.93 (0.91–0.94)	5.30 × 10−99	0.84	0.82	0.57	0.53
TCGA	HNSC	391	25.6	0.9 (0.88–0.92)	9.78 × 10−142	0.64	0.62	0.41	0.39
TCGA	UCEC	477	57.0	0.85 (0.83–0.88)	3.07 × 10−137	0.92	0.94	0.85	0.88
TCGA	BLCA	378	93.9	0.94 (0.93–0.95)	4.19 × 10−178	0.9	0.99	0.87	0.98
TCGA	STAD	371	60.1	0.92 (0.9–0.93)	7.89 × 10−152	0.77	0.82	0.75	0.81
CPTAC	BRCA	100	N/A	0.52 (0.36–0.65)	2.33 × 10−8	N/A	N/A	N/A	N/A
CPTAC	PAAD	139	22.3	0.85 (0.79–0.89)	1.65 × 10−39	0.65	0.63	0.39	0.33
CPTAC	HNSC	107	N/A	0.84 (0.77–0.89)	1.01 × 10−29	N/A	N/A	N/A	N/A
CPTAC	UCEC	99	26.3	0.83 (0.76–0.88)	2.02 × 10−26	0.83	0.79	0.76	0.62