RATIONALE AND OBJECTIVES: To evaluate an MRI radiomics-powered machine learning (ML) model's performance for the identification of deep myometrial invasion (DMI) in endometrial cancer (EC) patients and explore its clinical applicability. MATERIALS AND METHODS: Preoperative MRI scans of EC patients were retrospectively selected. Three radiologists performed whole-lesion segmentation on T2-weighted images for feature extraction. Feature robustness was tested before randomly splitting the population in training and test sets (80/20% proportion). A multistep feature selection was applied to the first, excluding noninformative, low variance features and redundant, highly-intercorrelated ones. A Random Forest wrapper was used to identify the most informative among the remaining. An ensemble of J48 decision trees was tuned and finalized in the training set using 10-fold cross-validation, and then assessed on the test set. A radiologist evaluated all MRI scans without and with the aid of ML to detect the presence of DMI. McNemars's test was employed to compare the two readings. RESULTS: Of the 54 patients included, 17 had DMI. In all, 1132 features were extracted. After feature selection, the Random Forest wrapper identified the three most informative which were used for ML training. The classifier reached an accuracy of 86% and 91% and areas under the Receiver Operating Characteristic curve of 0.92 and 0.94 in the cross-validation and final testing, respectively. The radiologist performance increased from 82% to 100% when using ML (p = 0.48). CONCLUSION: We proved the feasibility of a radiomics-powered ML model for DMI detection on MR T2-w images that might help radiologists to increase their performance.
RATIONALE AND OBJECTIVES: To evaluate an MRI radiomics-powered machine learning (ML) model's performance for the identification of deep myometrial invasion (DMI) in endometrial cancer (EC) patients and explore its clinical applicability. MATERIALS AND METHODS: Preoperative MRI scans of ECpatients were retrospectively selected. Three radiologists performed whole-lesion segmentation on T2-weighted images for feature extraction. Feature robustness was tested before randomly splitting the population in training and test sets (80/20% proportion). A multistep feature selection was applied to the first, excluding noninformative, low variance features and redundant, highly-intercorrelated ones. A Random Forest wrapper was used to identify the most informative among the remaining. An ensemble of J48 decision trees was tuned and finalized in the training set using 10-fold cross-validation, and then assessed on the test set. A radiologist evaluated all MRI scans without and with the aid of ML to detect the presence of DMI. McNemars's test was employed to compare the two readings. RESULTS: Of the 54 patients included, 17 had DMI. In all, 1132 features were extracted. After feature selection, the Random Forest wrapper identified the three most informative which were used for ML training. The classifier reached an accuracy of 86% and 91% and areas under the Receiver Operating Characteristic curve of 0.92 and 0.94 in the cross-validation and final testing, respectively. The radiologist performance increased from 82% to 100% when using ML (p = 0.48). CONCLUSION: We proved the feasibility of a radiomics-powered ML model for DMI detection on MR T2-w images that might help radiologists to increase their performance.