Abstract
3085
Introduction: Radiomics or computer-extracted imaging features have shown promise at predicting disease progression when applied to PET images. We evaluated the utility of prostate-specific membrane antigen (PSMA)-based 18F-DCFPyL (PyL) PET radiomics features in predicting disease progression in men with high-risk primary prostate cancer (PCa).
Methods: Patients with newly diagnosed high-risk PCa underwent baseline PyL pelvic PET/MRI (PET1) prior to 3 cycles of neoadjuvant chemohormonal therapy with docetaxel and androgen deprivation therapy. PyL pelvic PET/MRI (PET2) was obtained after neoadjuvant chemohormonal therapy and followed by radical prostatectomy. High-risk PCa eligibility was defined as: 1) extracapsular extension (cT3a) or seminal vesicle involvement (cT3b) or invasion of adjacent structures (cT4), serum PSA >20 ng/mL or a Gleason score of 8 to 10 and/or regional lymph node enlargement and 2) oligometastatic disease, defined as disseminated metastases beyond regional lymph nodes by standard-of-care imaging with three or less bone metastases but no visceral metastases. Patients underwent follow-up PyL PET/MRI imaging at one year after prostatectomy, or earlier if they met the criteria of prostate-specific antigen (PSA) progression, and subsequent clinical follow-up. Primary PCa lesions were assessed at both imaging time points (PET1 and PET2). Prostates were contoured based on oblique axial T2W MRI by an experienced genitourinary radiologist and transferred to PET images using MIM Encore software v6.8.7. To assess primary PCa heterogeneity, radiomics features were extracted from PET1 including the shape and size as well as higher-order features using the LIFEx software v7.1.0. To assess primary PCa therapy response, PET standardized updated value (SUV) parameters including SUVmax were assessed at PET1 and PET2. All PET features were correlated with time to PSA progression using univariable Cox regression. Least absolute shrinkage and selection operator (LASSO) regression was used to select key variables for multivariable Cox regression analysis. To determine a baseline false-discovery rate, the statistical analysis was repeated with sham data (i.e., random variables).
Results: Twenty-seven (mean age: 61.2 years, range: 44-71) were enrolled. Fifty-two radiomics features including nine conventional metabolic parameters were extracted and analyzed. Nine patients were disease-free, and 18 patients had PSA progression with a median time to progression of 209 days (range 63-587). Median clinical follow-up was 758 days (range 378-1140). Univariable analysis found 29 radiomics features including SUVmax in PET1 and PET2 to be significant predictors of time to progression. Three features including gray-level zone length matrix (GLZLM)_gray-level non-uniformity (GLNU) of PET1, skewness of PET1, and PET2 SUVmax were selected by the Lasso regression. Multivariable Cox regression analysis found GLZLM_GLNU of PET1 (odds ratio: 1.154, p = 0.005) and SUVmax of PET2 (odds ratio: 1.053, p = 0.017) to be significant predictors of progression (Table 1). In comparison, analysis of sham data found 5 significant variables in univariable regression and 1 significant variable in multivariable regression after LASSO selection. For the multivariable model, pseudo-R2 was 0.61 using real data vs 0.36 using sham data.
Conclusions: Radiomics features of PSMA-based PyL PET1 and SUVmax of PET2 may be useful biomarkers of disease progression in high-risk PCa patients undergoing neoadjuvant chemohormonal therapy. Further studies with larger cohorts are needed to confirm the clinical utility of these parameters.
Funding resources:
DOD PCRP Impact Award PC15053 (UW-Madison)
NIH/NCI 1P41EB024495-01 (JHU) - Service Project 2 (UW-Madison)