MRI‐based radiomic features for identifying recurrent prostate cancer after proton radiation therapy

Abstract

PurposeMagnetic Resonance Imaging (MRI) evaluation of recurrent prostate cancer (PCa) following proton beam therapy is challenging due to radiation‐induced tissue changes. This study aimed to evaluate MRI‐based radiomic features so as to identify the recurrent PCa after proton therapy.MethodsWe retrospectively studied 12 patients with biochemical recurrence (BCR) following proton therapy. Two experienced radiologists identified prostate lesions from multi‐parametric MRI (mpMRI) images post‐proton therapy and marked control regions of interest (ROIs) on the contralateral side of the prostate gland. A total of 210 radiomic features were extracted from lesions and control regions on the T2‐weighted (T2WI) and Apparent Diffusion Coefficient (ADC) image series. Recursive Feature Elimination with Cross‐Validation method (RFE‐CV) was used for feature selection. A Multilayer Perceptron (MLP) neural network was developed to classify three classes: cancerous, benign, and healthy tissue. The 12‐core biopsy results were used as the gold standard for the segmentations. The classifier performance was measured using specificity, sensitivity, the area under receiver operating characteristic curve (AUC), and other statistical indicators.ResultsBased on biopsy results, 10 lesions were identified as PCa recurrence while eight lesions were confirmed to be benign. Ten radiomic features (10/210) were selected to build the multi‐class classifier. The radiomics classifier gave an accuracy of 0.83 in identifying cancerous, benign, and healthy tissue with a sensitivity of 0.80 and specificity of 0.85. The model yielded an AUC of 0.87, 95% CI [0.72–1.00] in differentiating cancer from the benign and healthy tissues.ConclusionsOur proof‐of‐concept study demonstrates the potential of using radiomic features as part of the differential diagnosis of PCa on mpMRI following proton therapy. The results need to be validated in a larger cohort.

New Parametric 2D Curves for Modeling Prostate Shape in Magnetic Resonance Images

“…Popular methods for deep learning segmentation are based on convolutional neural networks, encoder-decoder and autoencoder models or generative adversarial networks [37]. More generally, artificial intelligence techniques are used in the medical field not only for segmentation but also for classification and prediction (radiomics) [38][39][40][41][42].…”