Session Item

The aim of this study was to compare prediction performance of biochemical failure for prostate cancer patients treated with low dose rate brachytherapy using Iodine-125 seeds between Machine Learning (ML) models using radiomic and dosiomic features.

Ninety-nine prostate cancer patients treated with Iodine-125 seed implantation between October 2009 and March 2013 were included in this study. Fourteen patients were observed with biochemical failure (BF) and 85 were not. CT scans with 2-mm slice thickness were examined one month after implantation and dose distribution (DD) was calculated on the CT slices using Variseed Ver. 9.0.2 software provided by Varian Medical Systems (Palo Alto, CA). Radiomic and dosiomic features within region-of-interest (ROI) of prostate were extracted from CT slices and DDs, and 8 kinds of their wavelet-transformed images and Laplacian of Gaussian (LoG) filtered images (Sigma = 1.0 and 1.5), using PyRadiomics Ver. 3.0.1. Obtained 1038 feature values were normalized with z-score. Pearson’s correlation coefficient with the threshold of 0.8 was used in order to remove highly correlated features, and then features which p-values of Mann Whitney’s U-test were less than 0.05 were employed in ML modeling. Those feature selection methods were implemented sequentially in each trial of leave-one-out cross-validation (LOOCV) flamework. ML methods used in this study were logistic regression with L1 regularization term (LRL1) and L2 term (LRL2), and random forest classifier (RFC), provided in Scikit-Learn (Ver. 0.23.1) ML library. Prediction models were validated in LOOCV framework and compared using area under the curve (AUC) values of the receiver operator characteristic analysis. Importance of features in RFC and normalized absolute weight values for features in logistic regression models were also obtained and discussed.

The AUC values of LRL1, LRL2, and RFC using radiomic features were 0.513, 0.624, and 0.645, respectively, and the ones using dosiomic features were 0.650, 0.922, and 0.697, respectively. Ninety-nine times of feature selection process and modeling trials were implemented in LOOCV modeling cycles, and 8 dosiomic features out of 1038 were selected more than 90 times. Analysis of normalized importance in RFC and absolute weight values in LRL1 and LRL2 for 8 dosiomic features shows that wavelet-LLH_NGTDM_Busyness and Strength shows higher importance in LRL1 and RFC whereas importance are widely distributed, including volume of prostate and Busyness of LoG filtered image, in the case of LRL2.

Dosiomic prediction models were predicting biochemical failure superior than radiomic models. Logistic regression model with L2 regularization using dosiomic features were resulting the AUC value of 0.922, which was the best in this study.