Prostate Cancer Nodal Staging: Using Deep Learning to Predict ⁶⁸Ga-PSMA-Positivity from CT Imaging Alone

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lymphatic spread is crucial for prostate cancer (PCa) treatment decisions.
• 68Ga-PSMA-PET/CT is accurate but limited by cost and availability.
• Computed tomography (CT) remains the standard for PCa staging.

Purpose of the Study:

• To evaluate if CNNs can determine lymph node status in PCa from CT images alone.
• To compare CNN performance against radiologists and random forest classifiers.
• To investigate the impact of training data balancing on CNN performance.

Main Methods:

• Trained three CNNs on 2616 lymph nodes from 549 patients with 68Ga-PSMA-PET/CT data.
• Used PET as the reference standard for lymph node status.
• Employed balanced training sets (infiltration, location, masked images) and evaluated on a separate test set.

Main Results:

• CNNs achieved an Area-Under-the-Curve (AUC) of 0.95 (status balanced) and 0.86 (location balanced, masked).
• CNN performance surpassed experienced radiologists (AUC 0.81).
• CNNs utilized anatomical context to improve predictions, learning infiltration probabilities of locations.

Conclusions:

• CNNs demonstrate potential as CT-based biomarkers for PCa lymph node metastases.
• AI can potentially provide a more accessible staging tool for PCa.
• Class balancing strategies significantly influence CNN performance in this application.