Prostate cancer is one of the most common cancers and leading causes of cancer-related death in men worldwide. Accurate staging -- identifying whether the cancer is confined to the prostate or has spread to lymph nodes or bones -- is critical for making effective treatment decisions.
PSMA PET-CT (prostate-specific membrane antigen positron emission tomography with computed tomography) has emerged as one of the most powerful tools for staging and restaging prostate cancer. PSMA is a protein that is significantly overexpressed on malignant prostate cells, making it a highly specific target for radiolabeled tracers that light up cancer sites on PET scans.
Compared to the conventional imaging standard of CT scan plus bone scan, PSMA PET-CT is more sensitive and more accurate at detecting both the primary tumor and distant metastases. The tracer used in this study, [18F]PSMA-1007, is one of several fluorine-18 labeled PSMA radiopharmaceuticals now available, offering practical advantages over gallium-68 labeled alternatives due to longer half-life and wider distribution potential.
Despite its advantages, PSMA PET-CT image interpretation relies on visual inspection by a nuclear medicine physician, which is time-consuming and subject to inter- and intra-observer variability. With PET-CT scan volumes rising rapidly and physician shortages widening, there is a pressing clinical need for automated AI tools that can detect and quantify prostate cancer lesions with consistent, physician-level accuracy.
The study enrolled 660 patients referred for clinically indicated [18F]PSMA-1007 PET-CT scans at Skåne University Hospital in Sweden between December 2019 and December 2020. Patients were included either for initial staging of high-risk prostate cancer or for detection of suspected disease recurrence after prior treatment, making the dataset clinically representative of the two most common PSMA PET-CT indications.
All patients received an injection of 4 MBq/kg of [18F]PSMA-1007 and underwent whole-body scanning from skull base to mid-thigh two hours later. The PET images were acquired at 2 minutes per bed position and reconstructed using a sophisticated algorithm that incorporates time-of-flight and point spread function modeling for high-quality image fidelity.
One experienced nuclear medicine physician with ten years of PET-CT experience manually segmented all suspected prostate tumors or local recurrences, lymph node metastases, and bone metastases across all 660 patients. These segmentations served as the ground truth labels for training the AI model. The 660 cases were split into 420 for training, 120 for validation, and 120 for independent testing.
To assess how the AI compared to human performance, six additional nuclear medicine physicians each independently segmented 40 cases from the 120-patient test set, producing two independent sets of expert readings for every patient in the test set. This multi-reader design enabled a rigorous comparison of AI performance against the natural variability between human experts.
The AI model is based on Unet3D, a three-dimensional convolutional neural network architecture designed for volumetric medical image segmentation. The network takes three simultaneous inputs for each patient scan: the CT image (normalized to a clinically relevant range of -800 to 800 Hounsfield units), the PET image (normalized over a standardized uptake range), and a multi-channel organ mask automatically derived from the CT image.
The organ mask identifies the anatomical location of key structures within the body. By providing this spatial context alongside raw CT and PET signal values, the network can learn not just that an area has high PSMA tracer uptake, but whether that uptake is occurring in an anatomically expected location for cancer versus a location associated with known non-specific uptake patterns.
The network outputs a voxel-level classification of each location in the body as one of four categories: prostate tumor or local recurrence, lymph node metastasis, bone metastasis, or background. All images were resampled to a standardized voxel size of 1.37 x 1.37 x 2.79 mm to ensure consistent spatial resolution across the training dataset.
A specialized adaptive sampling strategy was used during training to address the class imbalance problem inherent in whole-body oncology scans, where the vast majority of voxels are background. Initially, 50 percent of training patches were drawn from foreground lesion regions and 50 percent from background. After the first half of training, the sampling weights were dynamically updated based on where the model made the largest errors, directing subsequent training toward the most challenging regions of each scan.
The network was trained using categorical cross-entropy loss, but with foreground cancer voxels given 2.5 times the weight of background voxels. This weighting deliberately biases the model toward sensitivity over specificity, accepting a slightly higher false positive rate to minimize missed cancer lesions -- a clinically appropriate trade-off when the primary goal is detecting all possible sites of disease.
Training used the Adam optimizer with Nesterov momentum at an initial learning rate of 0.001. The learning rate was halved whenever validation loss plateaued for five consecutive epochs. Dropout layers with a 0.25 dropout rate and L2 weight regularization were included to prevent overfitting on the 420-scan training set.
Data augmentation was applied during training, including random scaling, rotation, and intensity shifts in both the PET and CT channels, improving the model's robustness to the natural variation in patient body habitus, scan positioning, and image acquisition parameters across different scans.
Beyond lesion detection, the model was also evaluated for its ability to measure tumor burden, a clinically meaningful quantity that captures the overall volume and metabolic activity of all cancer sites combined. Tumor burden was quantified as the total lesion volume (TLV, the sum of all detected cancer voxels) and total lesion uptake (TLU, TLV multiplied by mean standardized uptake value), separately for prostate, lymph node, and bone compartments.
On the 120-patient independent test set, the AI achieved an average sensitivity of 79 percent for detecting prostate tumor or local recurrence, 79 percent for lymph node metastases, and 62 percent for bone metastases. The nuclear medicine physicians achieved corresponding average sensitivities of 78, 78, and 59 percent respectively -- confirming that the AI model's performance was statistically comparable to experienced human readers across all three lesion types.
The highest AI sensitivity values were obtained when the segmentations from the physician who provided the training data were used as the reference standard, reaching 95 percent for prostate, 88 percent for lymph node, and 89 percent for bone lesions. This reflects the expected alignment between an AI model's outputs and the reading style of the physician on whose annotations it was trained.
The main difference between AI and human readers was in the false positive rate. For lymph node metastases, the AI produced an average of 2.8 false positive lesions per patient compared to 0.5 for human reader-versus-reader comparisons. For bone metastases, the AI produced 3.0 false positives per patient compared to 1.8 between readers. This higher false positive rate reflects the known challenge of [18F]PSMA-1007, which can show non-specific uptake in benign bone lesions, especially in the ribs and pelvis, at a rate found in at least 50 percent of patients scanned with digital PET-CT systems.
Compared to a recent published benchmark, the current AI model showed substantially fewer false positives per patient: 2.8 for lymph node metastases versus 19.5 per patient in the prior study, and 3.0 for bone lesions versus 8.2, despite achieving comparable sensitivity. This balance between sensitivity and false positive rate makes the current system more practical for clinical deployment.
For the clinically important task of tumor burden quantification, correlations between AI measurements and expert physician measurements were statistically significant across all lesion types. The Spearman correlation ranged from R = 0.53 for total lesion volume of lymph node metastases to R = 0.83 for total lesion uptake of prostate tumor or recurrence, with correlations for all six combinations (TLV and TLU across three lesion types) reaching significance at p less than 0.05.
Total lesion volume and total lesion uptake ranges measured by the AI fell within the ranges produced by the three human reader sets for prostate tumor and bone metastasis categories. For lymph node metastases, total lesion uptake by the AI was slightly outside the range of the three human readers, suggesting that the AI detected slightly more PSMA-avid lymph node tissue in aggregate than any individual physician reading.
Tumor burden quantification has direct prognostic significance in prostate cancer: multiple studies have shown that higher total lesion volume and total lesion uptake on PSMA PET-CT are associated with shorter overall survival in patients with advanced disease. An automated tool that can rapidly and consistently measure tumor burden could therefore provide prognostic information without the time and variability of manual measurements.
A patient example illustrated the practical value of the AI output: the system correctly identified and annotated prostate tumor locations in blue, lymph node metastases in pink, and bone metastases in yellow across the whole-body scan, with the corresponding automatically measured TLV values closely matching the manual physician measurements for each compartment.
This study demonstrated that a fully automated AI-based method for prostate cancer detection and tumor burden quantification in whole-body [18F]PSMA-1007 PET-CT is achievable with sensitivity matching experienced nuclear medicine physicians. The tool addresses three practical clinical problems simultaneously: reducing inter-reader variability, shortening interpretation time, and enabling objective and repeatable tumor burden assessment.
Inter-observer variability in PSMA PET-CT interpretation was confirmed to be substantial even among experienced readers following standardized reporting criteria. For bone lesions in particular, the wide range in sensitivity between different physician pairings (29 to 87 percent) underscores the challenge of distinguishing malignant uptake from the non-specific bone uptake that [18F]PSMA-1007 frequently produces in benign bone lesions.
A key contribution of this work is the decision to make the AI tool freely available to other researchers at recomia.org, the same cloud-based platform used for image annotation during the study. This open access approach enables independent validation across different institutions, different PET-CT scanners, and different patient populations -- the prospective multi-site validation that is necessary before clinical deployment.
Acknowledged limitations include the single-institution dataset and the use of expert visual segmentations rather than histopathological confirmation as ground truth. The model was trained and tested exclusively on [18F]PSMA-1007 and may not perform equivalently with other PSMA radiopharmaceuticals. Future work will focus on validation in patients with higher tumor burden and external validation at other centers using different imaging protocols.