Abstract
In PSMA-ligand PET/CT imaging, standardized evaluation frameworks and image-derived parameters are increasingly used to support prostate cancer staging. Clinical applicability remains challenging wherever manual measurements of numerous suspected lesions are required. Deep learning methods are promising for automated image analysis, typically requiring extensive expert-annotated image datasets to reach sufficient accuracy. We developed a deep learning method to support image-based staging, investigating the use of training information from two radiotracers. In 173 subjects imaged with 68Ga-PSMA-11 PET/CT, we trained and evaluated a convolutional neural network (CNN) to both classify sites of elevated tracer uptake as nonsuspicious or suspicious for cancer and assign them an anatomical location. We evaluated training strategies to leverage information from a larger dataset of 18F-FDG PET/CT images, including transfer learning and combined training encoding the tracer type as input. The algorithm showed good agreement with expert assessment for identification and anatomical location classification of suspicious uptake sites, demonstrating that training examples from a different radiotracer can improve performance when PSMA-ligand data is restricted.
Introduction
Prostate cancer staging relies heavily on PET/CT imaging using tracers like 68Ga-PSMA-11, which targets the prostate-specific membrane antigen. While highly accurate, the manual evaluation of these scans presents a significant clinical bottleneck. Radiologists must identify, measure, and categorize numerous potential lesions across the entire body, a process that is labor-intensive, subject to inter-observer variability, and prone to error.
Deep learning (DL) offers a potential solution for automation. However, DL models are notoriously "data-hungry," and high-quality, expert-annotated 68Ga-PSMA-11 datasets are relatively scarce compared to older, more common tracers like 18F-FDG. This study addresses the "small data" problem by investigating "dual-tracer learning"—leveraging knowledge gained from common FDG scans to improve prostate-specific PSMA imaging.
Methodology
The core of the proposed system is a multi-task Convolutional Neural Network (CNN) designed to perform two primary functions:
- Classification: Distinguishing between sites of elevated tracer uptake that are "suspicious" for cancer and those that are "nonsuspicious" (physiological).
- Anatomical Localization: Assigning the uptake site to a specific anatomical region, such as bone or lymph nodes.
The model processes 13 coronal PET/CT multiplanar reformations (MPRs) centered around a region of interest. To overcome data limitations, two dual-tracer strategies were evaluated: Transfer Learning (pre-training on FDG data) and Combined Training (training on both tracers simultaneously using a "tracer encoding" binary input).
Datasets and Annotations
The primary dataset consisted of 173 subjects imaged with 68Ga-PSMA-11 PET/CT, split into development (121) and test (52) sets. A secondary, larger dataset of 18F-FDG PET/CT images was used for dual-tracer learning. Expert nuclear medicine physicians provided over 5,500 annotations of high-uptake regions to serve as the ground truth.
Results and Discussion
The dual-tracer approach, specifically the Combined Training strategy, significantly outperformed models trained on PSMA data alone. The model achieved 80.4% average precision for identifying suspicious uptake sites and 77% accuracy for anatomical localization in the test set.
Cross-Tracer Synergy
A key finding was that FDG and PSMA scans share significant "visual logic," particularly regarding CT anatomy and the general appearance of suspicious versus physiological uptake. The performance boost was most pronounced in anatomical localization, suggesting the model uses the larger FDG dataset to build a robust internal map of human anatomy that transfers effectively to PSMA scans.
Clinical Staging Agreement
The model's outputs were used to automatically assign N (regional lymph node) and M (distant metastasis) stages according to the standardized PROMISE miTNM framework. The model showed high agreement with expert assessments:
- 81% agreement for regional lymph node involvement (N-stage).
- 77% agreement for distant metastases (M-stage).
Conclusion
This research demonstrates that automated staging using deep learning is a viable path forward for prostate cancer management. By leveraging dual-tracer learning, models can overcome the limitations of small datasets to provide high-level clinical insights. While future work is needed to validate the model in patients with very high tumor burdens and to move toward fully automated segmentation, these results represent a significant step toward efficient, standardized assessment of cancer stage and tumor burden in clinical practice.
Summary updated