TABLE 2.

Validation and test metrics of multi‐branch CLAM models based on features from domain‐specific (SimSiam) and pretrained encoders at multiple resolutions.

			Accuracy		Cross‐entropy		ROC AUC
Encoder type	Downscale	Magnification	Test	Valid.	Test	Valid.	Test.	Valid.
ResNet50 pretrained	2×	20×	0.86 ± 0.03	0.88 ± 0.02	0.36 ± 0.07	0.35 ± 0.01	0.96 ± 0.02	0.91 ± 0.02
	4×	10×	0.82 ± 0.07	0.9 ± 0.04	0.38 ± 0.06	0.35 ± 0.03	0.91 ± 0.02	0.9 ± 0.03
	8×	5×	0.75 ± 0.03	0.83 ± 0.01	0.82 ± 0.3	0.42 ± 0.02	0.86 ± 0.04	0.9 ± 0.01
ResNet101, SimSiam, SGD	2×	20×	0.85 ± 0.04	0.9 ± 0.01	0.43 ± 0.11	0.28 ± 0.02	0.9 ± 0.04	0.92 ± 0.03
	4×	10×	0.9 ± 0.01	0.97 ± 0.03	0.26 ± 0.05	0.09 ± 0.05	0.98 ± 0.0	0.99 ± 0.01
	8×	5×	0.98 ± 0.02	0.93 ± 0.0	0.14 ± 0.08	0.13 ± 0.04	0.99 ± 0.01	0.99 ± 0.0

Note: The best test metrics were achieved at 8× downscale using a domain‐specific ResNet101 (98% accuracy, highlighted in bold font). Mean and standard deviations refer to 5 independently trained CLAM models.