Machine learning-based analysis of [18F]DCFPyL PET radiomics for risk stratification in primary prostate cancer

• DOI: 10.1007/s00259-020-04971-z
• Publication Date: 2020-07-31T12:03:54Z
• Authors: Matthijs C. F. Cy...; Bernard H. E. Jansen; Tim van de Brug; Daniela E. Oprea-...; Elisabeth Pfaehler; Bart M. de Vries; Reindert J. A. va...; Otto S. Hoekstra; André N. Vis; Ronald Boellaard
• Keywords: Male; Prostate cancer; PSMA PET-CT; Radiomics; Prostatic Neoplasms; Risk Assessment; 3. Good health; Machine Learning; 03 medical and health sciences; 0302 clinical medicine; Positron Emission Tomography Computed Tomography; Machine learning; Humans; Original Article; Prospective Studies

Abstract Purpose Quantitative prostate-specific membrane antigen (PSMA) PET analysis may provide for non-invasive and objective risk stratification of primary prostate cancer (PCa) patients. We determined the ability machine learning-based quantitative [ 18 F]DCFPyL metrics to predict metastatic disease or high-risk pathological tumor features. Methods In a prospective cohort study, 76 patients with intermediate- PCa scheduled robot-assisted radical prostatectomy extended pelvic lymph node dissection underwent pre-operative PET-CT. Primary tumors were delineated using 50–70% peak isocontour thresholds on images without partial-volume correction (PVC). Four hundred eighty standardized radiomic features extracted per tumor. Random forest models trained involvement (LNI), presence any metastasis, Gleason score ≥ 8, extracapsular extension (ECE). For comparison, also standard (SUVs, volume, total PSMA uptake). Model performance was validated 50 times repeated 5-fold cross-validation yielding mean receiver-operator characteristic curve AUC. Results The radiomics-based learning predicted LNI (AUC 0.86 ± 0.15, p < 0.01), nodal distant metastasis 0.14, (0.81 0.16, ECE (0.76 0.12, 0.01). highest AUCs reached lower than those models. prediction, PVC higher delineation threshold improved model stability. Machine pre-processing methods had minor impact performance. Conclusion can in These findings indicate that expression detected is related both histopathology tendency. Multicenter external validation needed determine benefits radiomics versus clinical practice.