The role of the tumor peripheral microenvironment to establish prostate cancer invasiveness is gaining interest. Radiomics is a rapidly growing research field, however there are still many methodological challenges to guarantee robustness and reproducibility of the models. We aimed to verify the feasibility of a semi-automated segmentation strategy for periprostatic tissue on axial T2-weighted images from 30 magnetic resonance imaging scans, test stability of hand-crafted radiomics features to multiple segmentation and their potential value in identification of extracapsular tumor extension using a machine learning approach. 1274 radiomics features were extracted from each volume of interest, with less than half (40 %) resulting stable at the ICC analysis. The trained Naïve Bayesian model correctly classified 63 % of instances aggregating the cross-validation data (AUC = 0.68). Although the performance of our machine learning model did not reach optimal results, the proposed segmentation approach could represent a facilitator for future research in the field.
Semi-Automated Image Segmentation of Peri-Prostatic Tissue on MRI and Radiomics Features Stability: A Feasibility Study for Locally Advanced Prostate Cancer Detection
Cuocolo, Renato;
2022
Abstract
The role of the tumor peripheral microenvironment to establish prostate cancer invasiveness is gaining interest. Radiomics is a rapidly growing research field, however there are still many methodological challenges to guarantee robustness and reproducibility of the models. We aimed to verify the feasibility of a semi-automated segmentation strategy for periprostatic tissue on axial T2-weighted images from 30 magnetic resonance imaging scans, test stability of hand-crafted radiomics features to multiple segmentation and their potential value in identification of extracapsular tumor extension using a machine learning approach. 1274 radiomics features were extracted from each volume of interest, with less than half (40 %) resulting stable at the ICC analysis. The trained Naïve Bayesian model correctly classified 63 % of instances aggregating the cross-validation data (AUC = 0.68). Although the performance of our machine learning model did not reach optimal results, the proposed segmentation approach could represent a facilitator for future research in the field.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.