COMPSTAT 2022: Start Registration
View Submission - COMPSTAT2022
Title: ML pipeline for radiomics-based survival analysis on CT images of patients with hepatic CRC metastases Authors:  Anna Theresa Stueber - Ludwig-Maximilians-Universitaet (LMU) Muenchen (Germany) [presenting]
Stefan Coors - LMU (Germany)
Katharina Jeblick - LMU Munich (Germany)
Andreas Mittermeier - University Hospital, LMU Munich (Germany)
Osman Oecal - LMU Munich (Germany)
Balthasar Schachtner - University Hospital LMU Munich (Germany)
Philipp Wesp - LMU Munich (Germany)
Max Seidensticker - LMU Munich (Germany)
Michael Ingrisch - University Hospital, LMU Munich (Germany)
Abstract: Using a statistically rigorous approach to compare the prognostic performance of different machine learning (ML) configurations and feature sources (clinical data (cd), liver and tumor radiomics from CT images) for survival analysis in patients with hepatic metastases. Prospectively CT images and cd from 431 patients with hepatic metastases of colorectal carcinoma were analyzed. Liver and tumor metastases were segmented automatically (nnU-net) and 1218 radiomics features (rad-liver, rad-tumor) were calculated. A large-scale ML benchmark-pipeline for survival/risk prediction consisting of preprocessing, feature selection, dimensionality reduction (PCA), hyperparameter tuning and training of different models - elastic-net regression, random survival forest (RSF) and gradient boosting - was developed and evaluated via 10-fold cross-validation (CV) using the metric integrated Brier-score (IBS). Addressing dependency structures in the setup, a mixed-model approach was used to compare algorithm and data configurations. 60 ML pipeline configurations were evaluated, showing RSF performs constantly equal or better than the other two algorithms with best/lowest IBS value when tuned, without PCA using cd+rad-tumor features with mean IBS 0.167 (95\%-CI: 0.158; 0.176). We investigated our comprehensive benchmark pipeline via a mixed-model evaluation for the optimization of radiomics-based risk prediction. Optimal prognostic performance was achieved with a tuned RSF on rad-tumor with cd.