Automated detection of retroperitoneal metastatic lymph nodes of testicular germ cell tumors from MR images using convolutional neural networks.
Authors
Affiliations (4)
Affiliations (4)
- Radiology, HUS Diagnostic Center, University of Helsinki and Helsinki University Hospital, PO BOX 340, 00029 HUS, Finland. Electronic address: [email protected].
- Radiology, HUS Diagnostic Center, University of Helsinki and Helsinki University Hospital, PO BOX 340, 00029 HUS, Finland. Electronic address: [email protected].
- Radiology, HUS Diagnostic Center, University of Helsinki and Helsinki University Hospital, PO BOX 340, 00029 HUS, Finland. Electronic address: [email protected].
- Radiology, HUS Diagnostic Center, University of Helsinki and Helsinki University Hospital, PO BOX 340, 00029 HUS, Finland. Electronic address: [email protected].
Abstract
The aim of the study was to investigate the hypothesis that machine learning-based methods can be trained to detect retroperitoneal lymph node metastasis of testicular germ cell tumours (TGCTs) from magnetic resonance imaging (MRI) diffusion -weighted images (DWIs) and T2-weighted images. Fifty patients were included, 46 cases of metastasis and 4 controls (mean age 33 years, range 20 to 65 years) and a retrospective analysis was done. The metastases were manually segmented from DWI- and T2-weighted images individually. Both groups were individually analysed by convolutional neural networks (CNNs) to find out the capability of artificial intelligence (AI) to detect metastatic retroperitoneal lymph nodes of TGCT. CNNs detected retroperitoneal lymph node metastases with good accuracy on DWI and performed almost as well on T2-weighted images. In the DWI test group, sensitivity was 0.85 and specificity 0.79 compared to 0.73 and 0.97 for T2-weighted images, respectively. CNNs detected larger lesions better than small lesions, which was the most evident finding in the T2-weighted images. There was no significant difference in detection performance between seminoma and nonseminoma patients. CNN-based automated analysis supported the hypothesis of this study and demonstrated the ability to detect retroperitoneal metastases of TGCT. Further studies involving larger cohorts are warranted before clinical implementation.