Related Experiment Video
Updated: Jun 13, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Predicting Failure of Active Surveillance in Desmoid-Type Fibromatosis Using Radiomics: An International Multi-center
Stefanie N Hakkesteegt1, Douwe J Spaanderman2, Chiara Colombo3
1Department of Surgical Oncology and Gastrointestinal Surgery, Erasmus MC Cancer Institute, University Medical Center Rotterdam, Rotterdam, The Netherlands.
Background:
Active surveillance (AS) is the first-line approach for desmoid-type fibromatosis (DTF). However, 30 % of patients require active treatment. Identifying these patients will help upfront to define a personalized treatment approach. This study assessed whether radiomics can predict AS failure in patients with DTF.
Methods:
This multicenter study included data from the Netherlands (NL), Italy (ITA), and Canada (CAN). The study included patients with extra-abdominal DTF initially managed with AS and baseline MRI. Tumors were segmented using a minimally interactive deep-learning method, and radiomics features were extracted from T1-weighted (T1W) and T2-weighted (T2W) MRI scans. Prediction models to predict AS failure versus no failure were created using various machine-learning approaches. Both an internal cross-validation using all available data and an external leave-one-country-out cross-validation were used to assess model performance.
Results:
The cohort included 200 patients (72 NL, 62 ITA, 66 CAN), with AS failing for 26 % of the patients. Internal validation of the T1W+T2W imaging model resulted in an overall area under the curve (AUC) of 0.69 (95 % confidence interval [CI] 0.60-0.79). External validation resulted in an AUC of 0.58 (95 % CI 0.42-0.74) in the Dutch cohort, 0.76 (95 % CI 0.60-0.91) in the Italian cohort, and 0.77 (95 % CI 0.65-0.89) in the Canadian cohort. Adding clinical features did not improve the models' performance.
Conclusions:
Predicting AS failure with radiomics showed reasonable performance and generalized well to the Italian and Canadian cohorts. Pending improvements to the model or patient selection, the authors' model shows potential to better identify which DTF patients will benefit from AS and which will not.
