Automated analysis of FDG-PET/CT imaging to monitor heterogeneous disease response in metastatic castration-resistant prostate cancer.

Abstract

251 Background: The heterogeneity of individual sites of disease in prostate cancer is well recognized and increases over time as the therapies administered promote divergent evolution. In clinical practice, this is often depicted in radiology reports as a mixed response, wherein some lesions improve and others progress or emerge, for which individualizing management is challenging. The problem is exacerbated in patients with high volume disease by the inability to concisely and effectively assess overall disease status which may mask the presence of emerging resistance that could benefit from early intervention and/or a change in therapy. Methods: Thirty-one sets of serial baseline and on-treatment FDG-PET/CT images done for the clinical management of patients with progressing mCRPC were analyzed using TRAQinform IQ software (AIQ Solutions). Individual regions of interest (ROI) identified and tracked across imaging time-points were analyzed for a range of features. The univariate prognostic weight of each feature was assessed with Cox regression models. Imaging features from single timepoints, heterogeneity of response features, and PSA values/dynamics were input into the separate TRAQinform Profile software, which was calculated to predict either time on treatment or overall survival using 3-fold cross-validation of a random survival forest. Individual case reviews were performed on select patients including TRAQinform IQ analytics, PSA trends, radiology reports, and physician notes to evaluate the potential additive benefit of the TRAQinform IQ output. Results: In general, imaging features were more strongly correlated with overall survival than PSA dynamics. After iterative feature selection, only imaging features were selected for TRAQinform Profile scores. In the case of predicting OS, the most important features were baseline total lesion glycolysis (TLG), and the number of new/progressing lesions. Patients with high TRAQinform Profile scores had shorter median survival times than those with low scores (630 vs 1326 days, p<0.01, c-index = 0.744). Of particular value was the finding that the software could identify indicators of oligo-progression for which directed radiotherapy could be considered. Also detected were early indications of widespread disease progression represented as changes in total SUV uptake and TRAQinform Profile scores, which could indicate that the patient was no longer benefitting from treatment and a change in therapy was needed. Conclusions: In pilot analysis, the TRAQinform Profile scores generated from baseline and early on-treatment FDG-PET images had treatment-agnostic prognostic value in mCRPC. Further validation is on-going. The TRAQinform IQ software can aid in identifying the presence of resistant lesions in the setting of otherwise grossly stable disease as determined by the treating physician.