Integrating PET/CT, radiomics and clinical data: An advanced multi-modal approach for lymph node metastasis prediction in prostate cancer.

The involvement of lymph nodes critically affects patient outcomes in prostate cancer. While traditional risk models use factors like stage and PSA levels, the detection of lymph node involvement through modalities like PET targeting PSMA with Ga radiotracer plays a pivotal role in guiding treatments ranging from prostatectomy to pelvic radiotherapy. This study aims to create a deep learning model to predict lymph node involvement in intermediate to high-risk prostate cancer patients using Ga-PSMA PET/CT imagery, radiomics features, and various clinical parameters. For this study, Ga-PSMA PET/CT scans and corresponding clinical data from 229 prostate cancer patients were retrospectively collected. An artificial intelligence model, integrating PET/CT fusion images, clinical data, and radiomics features, was developed using a slice-wise feature extractor and MNASNet for spatial feature extraction. The model was trained on 181 cases and tested on 48 cases. To assess the model's performance, a reader study was conducted on a balanced subset of the test data with five radiation oncologists. Among the 229 intermediate to high-risk patients with localized prostate cancer evaluated, 67 (30%) had lymph node metastasis, while 162 were non-metastatic. The proposed AI model achieved a mean accuracy of 0.85±0.03 and an F1 score of 0.73±0.03. In the reader study, radiation oncologists' mean evaluations showed lower metrics (accuracy 0.71±0.08, F1 score 0.70±0.07), compared to the AI model's mean accuracy of 0.79±0.02 and F1 score of 0.76±0.02. Our findings demonstrate the potential benefits of the proposed model in the clinical setting, particularly in enhancing decision-making by doctors in scenarios with high variability between readers, such as lymph node metastasis prediction.