TH-C-141-03: A Machine Learning Algorithm-Based Risk Stratification of Prostate Gland Lesions Using Quantitative and Texture Features From Diffusion Weighted MRI

Abstract

Purpose: To propose a machine learning algorithm‐based risk stratification method of prostate gland lesions using quantitative and texture features derived from diffusion weighted MRI (DW‐MRI). Methods: 43 patients with biopsy‐proven prostate cancer underwent DW‐MRIs of the prostate using an optimized single‐shot echo planar imaging sequence at 3T. Quantitative apparent diffusion coefficient (ADC) maps were generated via a mono‐exponential fit using three b‐values (50, 400 and 800s/mm2). 43 lesions were identified in peripheral zone (pz) and 13 in central gland (cg). Statistical and textural parameters were extracted from these lesions. Patients were grouped into low, intermediate and high risk group based on NCCN risk scoring guideline. 16 quantitative parameters (mean, kurtosis, percentile etc) and 5 texture parameters (energy, entropy, correlation, contrast and homogeneity) were calculated for each patient. Artificial Neural Network (ANN) was applied to classify patients into one of the three risk groups based on their MRI parameters. A sub‐group of patients were used as ANN training samples, and the rest was used as test samples. Classification accuracy, precision, and sensitivity were reported. Results: Both statistical and texture parameters showed strong correlation to Gleason score for only PZ tumors and not CG tumors. Classification sensitivity performed using 5 texture parameters for all tumors, pz tumors and cg tumors was 100%, 98% and 91% respectively. Sensitivity using quantitative parameters only was 97%, 95% and 100%. Finally using all the parameters, sensitivity was 93%, 95% and 81% implying that adding texture information is not very sensitive for cg tumors as compared to pz tumors. The positive predictive value (PPV) was 100% for all the above cases. Conclusion: ANN is a reasonably accurate algorithm with high sensitivity and PPV to classify prostate MRIs into risk groups. Further analysis will be done on the minimum number of training points required for the proposed technique.