Utility of radiomic zones for risk classification and clinical outcome predictions using supervised machine learning during simultaneous 11 C-choline PET/MRI acquisition in prostate cancer patients

Abstract

Purpose: In most radiomic studies related to cancer research, the traditional tumor-centric view has predominated. In this retrospective study, we go beyond the single-tumor region and investigate the utility of proposed radiomic zones for risk classification and clinical outcome predictions using radiomic features extracted from ¹¹ C-choline positron emission tomography (PET) imaging and supervised machine learning in prostate tumors.

Materials and methods: Seventy-seven prostate tumors were selected and delineated. The prostate organ was divided into three radiomic zones, with zone-1 being the metabolic tumor zone, zone-2 the proximal peripheral tumor zone, and zone-3 the extended peripheral tumor zone. LIFEx was used for PET-radiomic feature extraction. Risk groups were created using Gleason scores (GS), prostate-specific antigen (PSA) levels, clinical TNM staging, and progression-free survival (PFS). Random forest (RF) and AdaBoost advanced machine learning algorithms were used for supervised machine learning. Accuracy, positive predictive value, area under the receiver operating characteristic curve (AreaROC), and other metrics were calculated for comparisons of predictive performance between zones.

Results: For the GS risk classification group, the accuracies of risk classification predictions were 71%, 71%, and 67% using RF and 65%, 64%, and 63% using AdaBoost for zones -1, -2, and -3, respectively. For the PSA group, the accuracies of risk classification predictions were 74%, 65%, and 64% using RF and 76%, 66%, and 67% using AdaBoost for zones -1, -2, and -3, respectively. For the TNM group, the accuracies of risk classification predictions were 68%, 76%, and 78% using RF and 66%, 75%, and 80% using AdaBoost for zones -1, -2, and -3, respectively. For the PFS group, the accuracies of clinical outcome predictions were 77%, 75%, and 83% using RF and 77%, 74%, and 83% using AdaBoost in zones -1, -2, and -3, respectively.

Conclusions: We proposed three radiomic zones with different standard uptake value characteristics and created four risk groups of prostate cancer patients for testing this idea. We showed that these radiomic zones have different predicting strengths in classifying risk groups and might allow us to identify a radiomic zone with higher accuracy for patient outcome prediction.

Keywords: 11C-choline PET; machine learning; prostate cancer; radiomic zones; radiomics.