Anatomically guided self-adapting deep neural network for clinically significant prostate cancer detection on bi-parametric MRI: a multi-center study

Insights into Imaging

Table 1 The demographics, clinical, and imaging characteristics of the PI-CAI training and in-house data set

Variables	PI-CAI training data	PI-CAI validation data	PI-CAI testing data	In-House testing data	In-house fine-tuning data
# of Patients	1476	100	1000	1002	200
# of Scans	1500	100	1000	1036	200
Age (years)	66 (61–70)	NA	NA	68 (59–73)	67 (58–70)
Prostate Specific Antigen (ng/mL)	8.5 (6–13)	NA	NA	9.2 (5–12)	8.2 (6–10)
# of different MRI scanners	5 Siemens, 2 Philips	6 Siemens, 3 Philips	6 Siemens, 3 Philips	6 Siemens, 2 GE	6 Siemens, 2 GE
# of Centers	3	3	3	9	9
PI-RADS Category of Positive MRI lesions 3 4 5	246 438 403	NA	NA	308 188 240	63 40 46
# of Scans with Benign or Indolent PCa	1075	NA	NA	NA	150
# of Scans csPCa the	425	NA	NA	288	50
# of csPCa Lesions	465	NA	NA	342	52

All continuous variables are presented with median and interquartile ranges
csPCa clinically significant prostate cancer, PI-CAI prostate imaging: cancer AI, PI-RADS Prostate Imaging-Reporting and Data System
*The PI-CAI training data were obtained from Radboud University Medical Center, Ziekenhuisgroep Twente, University Medical Center Groningen, Norwegian University of Science and Technology. In-house data were obtained from Acibadem Mehmet Ali Aydinlar University’s Maslak Hospital, Altunizade Hospital, Atakent Hospital, Adana Hospital, Taksim Hospital, Kozyatagi Hospital, Kocaeli Hospital, Bodrum Hospital, Eskisehir Hospital