Search

Highlights: What are the main findings? This multicenter retrospective study, encompassing two hospital sites for training and leveraging data from forty-seven different hospitals for testing, aimed to develop machine learning models utilising both clinical and radiomics features (individually and in combination) extracted from pre-treatment CT scans. The objective was to predict PFS and OS in patients with PDAC. Among the various models developed, the model based on the combination of clinical and CT-derived tumour radiomics features demonstrated superior performance for OS with a C-index of 0.72 in the testing dataset. Conversely, the model based solely on clinical features excelled in predicting PFS, achieving a C-index of 0.70 in the testing dataset. What is the implication of the main finding? The analysis further highlighted the high relevance of clinical features both for OS and PFS outcomes in patients with PDAC. Purpose. This multicenter retrospective study aims to identify reliable clinical and radiomic features to build machine learning models that predict progression-free survival (PFS) and overall survival (OS) in pancreatic ductal adenocarcinoma (PDAC) patients. Methods. Between 2010 and 2020 pre-treatment contrast-enhanced CT scans of 287 pathology-confirmed PDAC patients from two sites of the Hopital Universitaire de Bruxelles (HUB) and from 47 hospitals within the HUB network were retrospectively analysed. Demographic, clinical, and survival data were also collected. Gross tumour volume (GTV) and non-tumoral pancreas (RPV) were semi-manually segmented and radiomics features were extracted. Patients from two HUB sites comprised the training dataset, while those from the remaining 47 hospitals of the HUB network constituted the testing dataset. A three-step method was used for feature selection. Based on the GradientBoostingSurvivalAnalysis classifier, different machine learning models were trained and tested to predict OS and PFS. Model performances were assessed using the C-index and Kaplan–Meier curves. SHAP analysis was applied to allow for post hoc interpretability. Results. A total of 107 radiomics features were extracted from each of the GTV and RPV. Fourteen subgroups of features were selected: clinical, GTV, RPV, clinical & GTV, clinical & GTV & RPV, GTV-volume and RPV-volume both for OS and PFS. Subsequently, 14 Gradient Boosting Survival Analysis models were trained and tested. In the testing dataset, the clinical & GTV model demonstrated the highest performance for OS (C-index: 0.72) among all other models, while for PFS, the clinical model exhibited a superior performance (C-index: 0.70). Conclusions. An integrated approach, combining clinical and radiomics features, excels in predicting OS, whereas clinical features demonstrate strong performance in PFS prediction. [ABSTRACT FROM AUTHOR]

• Mean ADC values are robust across different scanners with imaging protocol standardisation. • Radiomic features of ADC map in a homogeneous phantom show substantial variation between MRI systems, even after ComBat harmonisation. • Resampling to a smaller voxel size provides more repeatable and reproducible radiomic features than resampling to a larger voxel size. The aim of this study was to perform a quantitative quality assurance of diffusion-weighted MRI to assess the variability of the mean apparent diffusion coefficient (ADC) and other radiomic features across the scanners involved in the REGINA trial. The NIST/QIBA diffusion phantom was acquired on six 3 T scanners from five centres with a rectum-specific diffusion protocol. All sequences were repeated in each scan session without moving the phantom from the table. Linear interpolation to two isotropic voxel spacing (0.9 and 4 mm) was performed as well as the ComBat feature harmonisation method between scanners. The absolute accuracy error was evaluated for the mean ADC. Repeatability and reproducibility within-subject coefficients of variation (wCV) were computed for 142 radiomic features. For the mean ADC, accuracy error ranged between 0.1 % and 8.5 %, repeatability was <1 % and reproducibility was <3 % for diffusivity range between 0.4 and 1.1x10-3mm2/s. For the other radiomic features, wCV was below 10 % for 24 % and 15 % features for repeatability with resampling 0.9 mm and 4 mm, respectively, and 13 % and 11 % feature for reproducibility. ComBat method could improve significantly the wCV compared to reproducibility without ComBat (p-value < 0.001) but variation was still high for most of the features. Our study provided the first investigation of feature selection for development of robust predictive models in the REGINA trial, demonstrating the added value of such a quality assurance process to select conventional and radiomic features in prospective multicentre trials. [ABSTRACT FROM AUTHOR]