Table 2.

PET uptake classification performance obtained with combined training of a convolutional neural network using ⁶⁸Ga-PSMA-11 PET/CT and ¹⁸F-FDG PET/CT scans, evaluated on a hold-out test dataset of ⁶⁸Ga-PSMA-11 PET/CT scans. Performance metrics determined by pooling findings of all subjects together are reported. Summary statistics for performance metrics determined at per-subject level are also reported. 95% confidence intervals obtained via bootstrap resampling at subject level are reported in brackets

Tracer	68Ga-PSMA-11
Classification output	Nonsuspicious vs. suspicious
Summary statistic	Pooled (CI)	Per-subject
		Average	Min	Q1	Median	Q3	Max
Performance metric
APa	80.4 (71.1, 87.8)	-	-	-	-	-	-
Recall	81.1 (70.6, 90.1)	85.2	11.1	73.8	100.0	100.0	100.0
PPVb	66.8 (60.3, 72.7)	65.5	0.0	50.0	68.3	100.0	100.0
True positives	159 (114, 209)	3.1	0	1	2	3	19
False positives	79 (58, 102)	1.5	0	0	1	2	8
False negatives	37 (18, 59)	0.7	0	0	0	1	8
Classification output	Anatomical location
Performance metric
Accuracysuspicious	77.0 (70.0, 83.4)	78.4	0.0	57.5	95.0	100.0	100.0
Accuracyall	94.4 (92.4, 96.1)	94.1	78.6	90.9	94.9	100.0	100.0

^aAverage precision

^bPositive predictive value