In-depth Analysis of PSMA PET/CT-based Multimodal Deep Learning Model for Predicting Lymph Node Metastases in Prostate Cancer Patients

Background

Prostate cancer (PCA) is one of the most common malignant tumors in men and a leading cause of cancer-related deaths. In clinically localized prostate cancer patients, extended pelvic lymph node dissection (EPLND) is widely regarded as the most accurate method for lymph node staging. However, the broad scope of this surgical procedure not only increases the risks of intraoperative and postoperative complications but also prolongs operation time and raises medical costs. Although EPLND’s role in evaluating lymph node invasion (LNI) remains controversial, its efficiency in predicting LNI makes it necessary for many patients to undergo the procedure.

Currently, clinical decision-making primarily relies on predictive models, such as the Memorial Sloan Kettering Cancer Center (MSKCC) and Briganti-2017 models, to determine whether patients should undergo EPLND. However, these clinical predictive models often have high false-positive rates, with almost 70% of patients being subjected to this invasive procedure without significant risks. Positron emission tomography (PET), as an emerging molecular imaging technology, combined with prostate-specific membrane antigen (PSMA), has shown higher specificity and sensitivity in lymph node evaluation. However, due to limitations of PSMA PET in detecting micrometastases, its value in clinical decision-making has yet to be fully realized.

To address this issue, the research team proposed a novel multimodal deep learning model integrating PSMA PET/CT imaging to optimize LNI prediction while reducing unnecessary EPLND procedures.

Source and Authors

This study was conducted by scientists from Xiangya Hospital, Central South University in China, the Department of Nuclear Medicine at Bern University Hospital in Switzerland, the Technical University of Munich in Germany, and Chongqing Jiaotong University in China. Leading authors include Qiaoke Ma, Bei Chen, Robert Seifert, Rui Zhou, among others. The study was published in the European Journal of Nuclear Medicine and Molecular Imaging, accepted in 2024, and is expected to be officially published in 2025.

Research Workflow

Study Subjects and Grouping

This retrospective study included 116 prostate cancer patients who underwent [68Ga]Ga-PSMA-617 PET/CT scans between April 2020 and September 2024. All patients underwent radical prostatectomy (RP) with EPLND during surgery and were divided into a training group (82 patients) and a testing group (34 patients) in a ratio of 7:3. Patients were excluded if they had prior treatment before PET scans, incomplete clinical or pathological data, or if the interval between PET scans and surgery exceeded one month.

Data Collection and Imaging Analysis

The study collected clinical, imaging, and pathological data for all patients. Clinical parameters included patient age, initial prostate-specific antigen (PSA) levels, International Society of Urological Pathology (ISUP) grade grouping from systematic biopsy, and clinical tumor staging (TNM staging) based on multiparametric magnetic resonance imaging (mpMRI) or PSMA PET/CT imaging. Notably, PSMA PET/CT images were systematically scored by two nuclear medicine physicians according to the PSMA-RADS v2.0 guidelines.

A key aspect of the study was the use of the Med3D deep learning model to extract imaging features. Pretrained on large-scale 3D medical imaging datasets, Med3D was used to extract high-dimensional features from PSMA PET and CT images with robust adaptability and transferability.

Development of the Deep Learning Model

The study developed a multimodal deep learning predictive model based on a multi-kernel support vector machine (SVM) algorithm. The model inputs included PSMA PET/CT deep learning features, SUVmax (maximum standardized uptake value), and clinical parameters of the patients. The development process included:

1. Data Preprocessing: PET and CT images were normalized and cropped to enhance data quality.
2. Feature Extraction: High-order imaging features were extracted from the entire prostate and lesion volumes.
3. Model Training and Cross-Validation: The model was optimized through leave-one-out cross-validation and further validated with 5–5 and 10–10 stratified cross-validation to ensure stability.
4. Model Evaluation: The model’s performance was compared with traditional predictive models and visual evaluations of the PET images in the testing set.

Data Analysis Methods

The experimental results were analyzed using Receiver Operating Characteristic (ROC) curves to calculate the area under the curve (AUC). Calibration curves and Decision Curve Analysis (DCA) were used to evaluate prediction accuracy and the clinical application value of the model.

Results Analysis

Performance Comparison of Models

The proposed multimodal model achieved an AUC of 0.89 (95% CI: 0.81-0.97) in the training set, outperforming both PSMA PET visual evaluations (AUC 0.82) and the conventional MSKCC and Briganti-2017 models (AUC 0.75 and 0.73, respectively). The model demonstrated a sensitivity of 71% and a specificity of 97%, with high net benefit observed in both calibration curves and decision curve analysis.

In the testing set, the multimodal model again displayed stable and robust performance (AUC 0.85, CI: 0.69-1.0). Using a risk threshold of 31%, the multimodal model avoided approximately 50% of unnecessary EPLND procedures, while missing fewer than 10% of positive LNI cases.

Data Support

Compared with LNI-negative patients, LNI-positive patients exhibited significantly higher PSA levels (31.3 vs. 17.4 ng/mL, p=0.015) and SUVmax (16.7 vs. 13.7, p=0.022). Calibration tests using the Hosmer-Lemeshow test yielded p-values greater than 0.05, indicating high concordance between predicted and observed values.

Conclusions and Significance

This study is the first to propose a multimodal model combining PET imaging-based artificial intelligence assessments with traditional clinical parameters, demonstrating superior performance over conventional predictive tools for prostate cancer lymph node risk prediction. The results indicate that comprehensive evaluation covering whole-prostate deep learning features can significantly reduce unnecessary surgical procedures while lowering the risk of missed diagnoses.

Key Highlights

1. Innovative Approach: Introduction of Med3D deep learning features across the whole prostate in PSMA PET/CT imaging for the first time.
2. Clinical Value: Improved predictive accuracy and reading consistency while reducing both false-positive and false-negative risks.
3. Scalability: The model framework shows generalizability and can be enhanced using different imaging techniques and large-scale patient datasets in the future.

Despite limitations such as a small sample size and single-center design, this study provides a foundation for future cross-center external validations and more extensive research on molecular imaging.