1797P - Application of novel machine learning model in [68Ga] Ga-PSMA-11 PET/CT: Predicting survival in oligometastatic prostate cancer patients

Background

Biochemical recurrence is estimated to occur in ≥ 25% of patients with prostate cancer (PC) following primary curative therapy. Machine learning models are being developed for lesion detection and tracking to provide a comprehensive view of disease burden, allowing clinicians to quantify and predict effectiveness of treatment for individual lesions. This study applied novel AI-assisted technology to automatically extract features from [⁶⁸Ga] Ga-PSMA-11 (PSMA) PET/CT images that correlate with treatment intervention and survival data to create a scoring system.

Methods

Between 2015 and 2016, 185 men with oligometastatic PC had a baseline and follow-up PSMA PET/CT scan (at ∼6-months) whilst treated per standard clinical care. Lesions were quantified and matched between timepoints using AIQ Solutions technology. Imaging features were extracted from each patient, including change in basic features (SUV_max, SUV_mean, and number of lesions at baseline), and heterogeneity features (intrapatient heterogeneity of disease and response). Univariate predictive power of overall survival (OS) prediction of each measure was determined using Cox regression models. An AI approach was evaluated to predict OS using 5-fold cross-validation of a random survival forest. Model performance was evaluated using the c-index.

Results

The top univariate predictors of survival were all heterogeneity features, proportion of lesions increasing (c-index=0.62), number of stable lesions (0.62), number of decreasing lesions (0.60), and number of new lesions (0.59). In an individual scan, the proportion of increasing lesions >29% correlated with poorer progression. The AI model was able to predict responders vs suboptimal responders based on whether they had a treatment intervention or observation alone (35%) (c-index=0.83 in both cases).

Conclusions

This study demonstrates that an AI-assisted lesional response analysis can help predict response and prognosis of oligometastatic PC patients. These results support further studies to validate these findings in a prospective cohort.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.