Session Item

To develop a machine learning (ML) model for early detection of brain metastases based on diffusion imaging radiomics.

Diffusion weighted MRI from 40 patients previously treated at our institution were retrospectively analyzed. Clinical target volume contours from 193 metastases were extracted from radiosurgery planning CTs and rigidly registered to corresponding Gd-T1 MRI and Apparent Diffusion Coefficient (ADC) maps. Control volumes were generated using contralateral contours located in healthy brain tissue to enable ML binary classification.

The ML input dataset consisted of: 1) ADC-based radiomic features calculated within target volumes using Pyradiomics, 2) linear slopes and intercepts of each radiomic feature calculated using timepoints before the metastasis manifested on conventional Gd-T1, 3) primary cancer site and 4) anatomical target volume location data, identified by registering images to the MNI152 T1 dataset and applying standard cortical and subcortical atlases.

Correlation analysis was performed and any features with >95% Pearson correlation were excluded. The dataset was divided into training and validation sets using an 80/20 split with stratification and scaled using Scikit-Learn’s StandardScaler. Five classification algorithms (SVM: Support Vector Machine, RF: Random Forest, MLP: Multi-layer Perceptron, ADA: AdaBoost, XGB: XGBoost) performed supervised learning using a 10-fold cross validation (CV) training set, with data labeled as either ‘control’ or ‘metastasis’. Grid search was used to tune hyperparameters for each algorithm (CV = 10), optimizing towards classifier balanced accuracy score. Receiver-operator curve area (AUC) scores were calculated along with accuracy, recall, and precision.

ML algorithm performance is summarized in Table 1. Gradient boosting-based algorithms XGBoost and AdaBoost showed superior accuracy (XGB: 0.764 ± 0.053 and 0.790 ± 0.099, ADA: 0.746 ± 0.050 and 0.880 ± 0.074) for both training and validation sets, respectively. SVM, RF and MLP all performed lower during training but remained comparable to other models when tested against the validation set. 

Twenty of the 25 most important ADA features all involved the change in radiomic feature values within clinical target volumes. Twenty-three of the 25 features were derived using wavelet filtered or edge enhanced ADC images.

Gradient-boosting based ML algorithms showed encouraging results in differentiating healthy brain tissue from metastases when using diffusion images prior to the lesion detection on conventional T1 MRI. ADA and XGB performed better than RF and SVM-based models when detecting malignant tissue based on changes in diffusion weighted imaging radiomic features. Future work will use these findings to further refine the model by adding more patients and test cases, improving classification accuracy, and increasing the model’s overall clinical applicability.