Detection of prostate cancer using machine learning with PSMA-PET/CT and multiparametric MRI

Abstract

Prostate Cancer (PCa) is one of the most common cancers in men globally. Recently, approaches to personalise treatment for individual patients such as Biologically targeted Radiotherapy (BiRT) have been proposed. The BiRT approach aims to customise radiation dose distributions according to the biological heterogeneity of a patient’s tumour derived from imaging. Recent studies have shown that using Prostate Specific Membrane Antigen (PSMA) Positron Emission Tomography/ Computed Tomography (PET/CT) with multiparametric Magnetic Resonance Imaging (mpMRI) can improve the detection of PCa, but few studies have analysed the relationship between the imaging modalities. Hence, this study aimed to quantify relationships between PSMA-PET and mpMRI parameters, and investigate the performance of radiomics-based Machine Learning (ML) models using each modality to predict tumour location and grade. To achieve this, imaging data from 15 PCa patients were co-registered using an established registration framework, and voxel-wise correlation analysis was conducted between MRI parameters and PET SUV for the tumour regions and compared with data of eight patients from a previous study. Radiomics Features (RFs) were extracted from PSMA-PET and mpMRI images of all 23 patients to construct ML models to discriminate tumour voxels, and high and low-grade tumour voxels. The classification algorithms, Support Vector Machine (SVC) and Random Forest (RFC) were compared, built using five-fold cross-validation. Correlation analysis showed the Dynamic Contrast-Enhanced (DCE) MRI parameters: transfer constant (Ktrans), initial area under the contrast agent concentration curve for the first 60s after injection (iAUGC60), Initial Rate of Enhancement (IRE) and Maximum Enhancement (ME) showed positive correlations with PSMA-PET SUV, while Time to Peak (TTP) showed negative correlations with PSMA-PET SUV. The RFC tumour detection model performed best when using RFs from both PSMA-PET and mpMRI rather than each modality on its own yielding accuracy, sensitivity, specificity and area under the curve (AUC) of 0.804, 0.738, 0.809 and 0.841, respectively. The tumour grade prediction RFC model yielded balanced accuracy of high-grade and low-grade tumours of 0.500 to 0.882 and 0.494 to 0.861. This study provides insights into the benefits of using PSMA-PET with mpMRI for BiRT treatment planning in the future.