TABLE V.

Melanoma Live Cell Image Compression and Classification, and Batch Effect Contamination of Latent Representations

	Seen batches		Unseen batches		Latent batch clustering
Model	MSE	AUROC (95% CI)	MSE	AUROC (95% CI)	DB	CH
Conventional AEC	0.0019	0.817 (0.812 – 0.822)	0.0024	0.773 (0.764 – 0.781)	8.885	545.9
DA-AEC	0.0018	0.777 (0.771 – 0.783)	0.0024	0.759 (0.750 – 0.768)	43.009	20.4
ARMED-AEC	0.0012	0.869 (0.865 – 0.874)	0.0024	0.789 (0.781 – 0.798)	43.009	20.4
w/o Adv.	0.0012	0.876 (0.872 – 0.881)	0.0024	0.791 (0.782 – 0.799)	8.885	545.9
randomized Z	0.0018	0.732 (0.726 – 0.738)	0.0024	0.712 (0.702 – 0.721)

AEC: autoencoder-classifier; DA: domain adversarial; Adv.: adversary; MSE: mean squared error between original and reconstructed images; AUROC: area under receiver operating characteristic curve for phenotype classification; CI: confidence interval; DB: Davies-Bouldin score, lower values indicate stronger clustering; CH: Calinski-Harabasz score, higher values indicate stronger clustering

Confidence intervals were computed with DeLong’s method. The best results for each metric are bolded.