Correlating 68Ga-PSMA PET/CT Imaging Features with PSA Variation for Castration-Resistant Prostate Cancer Patients
Résumé
Aim/Introduction: Prostate cancer is a major health concern
worldwide, and the treatment of advanced prostate cancer is still
challenging. Radiopharmaceutical therapy with 177Lu-PSMA has
emerged as a promising approach to address this issue. PET/CT
imaging is commonly used to determine the extent of disease
and the correlation between PET/CT imaging features and serum
PSA levels has been shown to predict treatment outcomes [1].
In this work, we analyzed some correlations between imaging
features and PSA levels to highlight the importance of PET/
CT imaging in predicting the response to 177Lu-PSMA therapy.
Materials and Methods: We conducted a retrospective analysis
of 45 male patients who received 177Lu-PSMA treatment, all
patients had a 68Ga-PSMA PET/CT imaging. The images were
pre-processed using a deep learning automatic segmentation
(TotalSegmentator [2]); the physiological organs uptake, obtained
by expanding 68Ga-PSMA-avid known OARs (liver, kidneys, spleen,
stomach, pancreas ڮ(, was set to zero to obtain one single
whole-body tumor region (WBTR) encompassing all potential
tumors without physiological uptake. For each patient, the
biological PSA (initial and fnal) was retrieved, and we studied the
correlation between the PET radiomics and the PSA values using
Spearman’s rank correlation, a Multi-Layer Perceptron with three
hidden layers, and a Random Forest. We compared the results
to a standard approach where the WBTRstandard is obtained by
a threshold set as the kidney’s mean SUV. Results: A signifcant
but weak correlation was found between the WBTR volume and
the PSA variation. The correlation coefcient was -0.41 (p-value:
0.01), while for the standard approach it was -0.34 (p-value: 0.02).
Moreover, we evaluated the predictive performance of MLP and
Random Forest models for PSA variation. We obtained a mean
absolute error of 163% for MLP and 563% for Random Forest. The
standard approach had an error of 149% for MLP and 558% for
Random Forest. Conclusion: This preliminary study only showed
the presence of a weak correlation between the WBTR volume
and the PSA variation. The use of simple statistics is not enough
to establish a signifcant improvement. Further research through
advanced machine learning techniques is necessary to explore
the potential of this method. References: [1] Moazemi, May 2021,
Annals of Translational Medicine[2] Wasserthal, 2022, arXiv
Origine | Fichiers produits par l'(les) auteur(s) |
---|