Multicentric 68Ga-PSMA PET radiomics for treatment response assessment of 177Lu-PSMA-617 radioligand therapy in patients with metastatic castration-resistant prostate cancer

Abstract

Objective: The treatment with ¹⁷⁷Lutetium PSMA (¹⁷⁷Lu-PSMA) in patients with metastatic castration-resistant prostate cancer (mCRPC) has recently been approved by the FDA and EMA. Since treatment success is highly variable between patients, the prediction of treatment response and identification of short- and long-term survivors after treatment could help tailor mCRPC diagnosis and treatment accordingly. The aim of this study is to investigate the value of radiomic parameters extracted from pretreatment ⁶⁸Ga-PSMA PET images for the prediction of treatment response.

Methods: A total of 45 mCRPC patients treated with ¹⁷⁷Lu-PSMA-617 from two university hospital centers were retrospectively reviewed for this study. Radiomic features were extracted from the volumetric segmentations of metastases in the bone. A random forest model was trained and validated to predict treatment response based on age and conventionally used PET parameters, radiomic features and combinations thereof. Further, overall survival was predicted by using the identified radiomic signature and compared to a Cox regression model based on age and PET parameters.

Results: The machine learning model based on a combined radiomic signature of three features and patient age achieved an AUC of 0.82 in 5-fold cross-validation and outperformed models based on age and PET parameters or radiomic features (AUC, 0.75 and 0.76, respectively). A Cox regression model based on this radiomic signature showed the best performance to predict overall survival (C-index, 0.67).

Conclusion: Our results demonstrate that a machine learning model to predict response to ¹⁷⁷Lu-PSMA treatment based on a combination of radiomics and patient age outperforms a model based on age and PET parameters. Moreover, the identified radiomic signature based on pretreatment ⁶⁸Ga-PSMA PET images might be able to identify patients with an improved outcome and serve as a supportive tool in clinical decision making.

Keywords: PET/CT; artificial intelligence; machine learning; metastases; prognosis.

Figure 1

Feature distribution for age and PET parameters. SUV, standardized uptake value; MTV, metabolic tumor volume.

Figure 2

Receiver operating characteristic curves for response prediction models. AUC, area under the receiver operating characteristic curve; CV, cross-validation; FPR, false positive rate; SD, standard deviation; TPR, true positive rate.

Figure 3

Distribution of the radiomic features and patient's age between responders and non-responders. GLCM, gray-level co-occurrence matrix.

Figure 4

Kaplan-Meier survival curves for risk prediction based on multiple Cox regression models for the training data (A, top row) and the test data (B, bottom row).

Figure 5

Potential clinical implementation of the workflow combining the response assessment and prognosis estimation, as well as representative PET examples for lesions identified as high survival and low survival probability.