Utilizing nnU-Net With 68Ga-PSMA-11, 18F-DCFPyL PET/CT for Metastatic Prostate Cancer Staging

Radioactive tracer drugs for the staging of prostate cancer have demonstrated good performance and high accuracy, but they are also associated with pitfalls, including stage-determinant subcentimeter lesions that require close assessment by physicians. Utilizing deep learning methods such as nnU-Net for the identification of cancer sites and staging has shown promise in prostate-specific membrane antigen (PSMA)-ligand positron emission tomography (PET)/computed tomography (CT).

A team of researchers examined the use of a convolutional neural network-based method in conjunction with ⁶⁸Ga-PSMA-11 and ¹⁸F-DCFPyL PET/CT to segment and localize sites that may be suspicious for prostate cancer for nodal and metastatic staging according to the molecular imaging tumor, node, metastasis (miTNM) framework.

The research team utilized nnU-Net, a deep learning-based segmentation method that automatically adapts to a given dataset to configure itself for new tasks, to segment suspected cancerous regions. Through the nnU-Net framework, an encoder-decoder network employing 3-dimensional convolutions was trained to segment regions that were suspected of cancer.

To ensure the network’s efficacy, an algorithm was developed to analyze its performance when training a separate network for each radiotracer compared with training a single network for both radiotracers. The best-performing configuration was then trained on the full development set and evaluated on the test set. A fully automated estimation of the prostate cancer stage was determined following the miTNM framework and compared with the stage assigned by the research team for both nodal and metastatic stages.

A total of 193 patients were referred to ¹⁸F-DCFPyL PET/CT for biochemical recurrence, while 173 patients were referred to ⁶⁸Ga-PSMA-11 for all indications of prostate cancer. Results were analyzed, and suspected regions for prostate cancer were segmented and assigned an anatomical location classification. Researchers discovered 450 suspicious findings from the ¹⁸F-DCFPyL PET/CT scans and 1057 suspicious findings from the ⁶⁸Ga-PSMA-11 scans.

The findings that correlated to local recurrence, lymph nodes, bone, and other sites for ¹⁸F-DCFPyL were 39, 280, 102, and 29, respectively. The findings that correlated to local recurrence, lymph nodes, bone, and other sites for ⁶⁸Ga-PSMA-11 were 30, 445, 529, and 53, respectively.

Training a single nnU-Net network for both radiotracers resulted in higher overall performance for detecting suspicious sites compared with training separate networks for each radiotracer. The assigned miN and miM stages corresponded to the stage that was predicted via deep learning in 65% and 73% of cases for ¹⁸F-DCFPyL and 79% and 75% of cases for ⁶⁸Ga-PSMA-11, respectively.

Patients who were screened with ⁶⁸Ga-PSMA-11 were found to have a higher disease burden, possibly demonstrating underlying observed differences in staging performance.

Deep learning methods can be used as effective identifiers of regions that may be suspicious for cancer. These methods have been shown to determine miTNM nodal and metastatic staging in ¹⁸F-DCFPyL PET/CT and ⁶⁸Ga-PSMA-11 images with moderate to good agreement.