Abstract
Prostate-specific membrane antigen (PSMA) positron emission tomography (PET)/computed tomography (CT) is a molecular and functional imaging modality with better restaging accuracy over conventional imaging for detecting prostate cancer in men suspected of lymph node (LN) progression after definitive therapy. However, the availability of PSMA PET/CT is limited in both low-resource settings and for repeating imaging surveillance. In contrast, CT is widely available, cost-effective, and routinely performed as part of patient follow-up or radiotherapy workflow. Compared with the molecular activities, the morphological and texture changes of subclinical LNs in CT are subtle, making manual detection of positive LNs infeasible. Instead, we harness the power of artificial intelligence for automated LN detection on CT. We examined <sup>68</sup>Ga-PSMA-11 PET/CT images from 88 patients (including 739 PSMA PET/CT-positive pelvic LNs) who experienced a biochemical recurrence after radical prostatectomy and presented for salvage radiotherapy with prostate-specific antigen < 1 ng/mL. Scans were divided into a training set (nPatient = 52, nNode = 400), a validation set (nPatient = 18, nNode = 143), and a test set (nPatient = 18, nNodes = 196). Using PSMA PET/CT as the ground truth and consensus pelvic LN clinical target volumes as search regions, a 2.5-dimensional (2.5D) Mask R-CNN based object detection framework was trained. The entire framework contained whole slice imaging pretraining, masked-out region fine-tuning, prediction post-processing, and "window bagging". Following an additional preprocessing step-pelvic LN clinical target volume extraction, our pipeline located positive pelvic LNs solely based on CT scans. Our pipeline could achieve a sensitivity of 83.351%, specificity of 58.621% out of 196 positive pelvic LNs from 18 patients in the test set, of which most of the false positives can be post-removable by radiologists. Our tool may aid CT-based detection of pelvic LN metastasis and triage patients most unlikely to benefit from the PSMA PET/CT scan.
Introduction
Identifying metastatic pelvic lymph nodes is critical for planning salvage radiotherapy in prostate cancer patients experiencing biochemical recurrence. While PSMA PET/CT is the current gold standard, its high cost and limited availability often restrict its use. Standard CT imaging is ubiquitous but suffers from low sensitivity for subclinical metastases; the morphological and texture changes are so subtle that they are virtually invisible to the human eye. This study explores whether a deep learning pipeline can act as a "super-observer," identifying PET-positive signatures within standard CT images to bridge this diagnostic gap.
Methodology
The researchers developed a multi-stage 2.5D object detection pipeline based on the Mask R-CNN architecture.
Data Preprocessing and 2.5D Representation
To provide the spatial context necessary to distinguish spherical lymph nodes from tubular blood vessels, the model processes three adjacent CT slices (central, superior, and inferior) concatenated as channels. Raw DICOM values are converted to Hounsfield Units (HU), and the search area is restricted to "Clinical Target Volumes" (CTVs)—regions where lymph nodes are most likely to reside—to mitigate class imbalance.
Model Architecture and "Window Bagging"
The framework employs a two-stage training strategy. First, a WSI-Mask R-CNN is pretrained on whole-slice images to learn general pelvic anatomy. This is followed by a Regional Mask R-CNN fine-tuned on masked-out regions to sharpen focus on minute lesions.
A key innovation is "Window Bagging," an ensemble approach where multiple models are trained on different HU ranges (quantile windowing). These models then "vote" on the final detection, significantly reducing false positives caused by varying tissue appearances.
Datasets
The study utilized a multi-institutional dataset of 88 patients from four institutions across Germany and the USA. A total of 739 PSMA PET/CT-positive pelvic lymph nodes served as the ground truth. The data was split into training (52 patients, 400 nodes), validation (18 patients, 143 nodes), and test sets (18 patients, 196 nodes).
Key Results
The final pipeline demonstrated strong performance on the independent test set:
- Sensitivity: 83.351%
- Specificity: 58.621%
- AUC (Area Under the Curve): 90.03%
Ablation studies showed that the inclusion of WSI pretraining and window bagging progressively improved precision by approximately 5% at each step compared to standard ImageNet-pretrained models. While specificity remains a challenge, the authors noted that many false positives (such as bowel segments or vascular structures) are easily identifiable and removable by a radiologist.
Discussion and Future Work
This AI-driven approach demonstrates that deep learning can extract features from CT scans that human experts cannot. By limiting search regions to pelvic LN CTVs and using ensemble voting via windowing, the model achieves high sensitivity for detecting metastatic nodes.
However, limitations remain. The dataset size is relatively small for deep learning, and the pipeline currently relies on manually contoured CTVs. Future iterations aim to integrate an automated pelvic LN segmentation network to create a fully autonomous workflow. Currently, the tool is best suited as a triage mechanism to identify patients most likely to benefit from a confirmatory PSMA PET/CT scan.
Summary updated