Machine Learning-Based Whole Gland Radiomics Analysis for Prostate Cancer Classification

Prostate cancer (PCa) is the second more common cancer type in men, and the need for a non-invasive PCa diagnostic and stratification tool can now be met via quantified analysis of Multiparametric MRI. Radiomics analysis can quantify imaging information on a Region of Interest (ROI), not visible by clinicians and thus improve cancer detection. However, the delineation of cancer lesions is both challenging and time consuming. In this paper, we consider as area in focus the whole prostate gland as segmented by high performance deep learning algorithms, calculate whole gland radiomics features and employ them to develop machine learning models for the identification of clinically significant Pca (csPCa). T2 weighted and diffusion weighted images from 1500 patients were used for model development and validation. The radiomic features from the prostate gland’s core, periphery and whole volume were extracted. The radiomic features were harmonized against vendor batch effect using statistical methods and a robust repeated elastic net method was employed for feature selection. Clinical data were also considered to improve model performance. Finally, Support Vector Machines and Balanced Random Forest (BRF) models were developed, and the validation step was repeated 5 times by splitting the whole dataset in 1:3 ratio in each iteration. A balanced accuracy of 79% was achieved for the SVM using radiomics from the whole prostate gland. The final model’s good performance highlights the value of whole gland radiomics for the identification of csPCa.