Digital Health Innovation and Informatics

PV 02 - Poster Viewing Q&A - Session 2

MO_19_2320 - A Machine Learning Model using MRI-based Radiomic Features to Predict Primary Site for Brain Metastases

Monday, September 16
10:45 AM - 12:00 PM
Location: ASTRO Innovation Hub

A Machine Learning Model using MRI-based Radiomic Features to Predict Primary Site for Brain Metastases
T. H. Felefly1, S. Achkar1, T. Lteif2, F. Azoury1, C. Khoury1, R. Sayah1, J. Barouky1, N. Farah1, D. Nehme Nasr1, and E. Nasr1; 1Hôtel Dieu de France Hospital, Beirut, Lebanon, 2Saint Joseph University, Beirut, Lebanon

Purpose/Objective(s): The most common primary sites for brain metastases are lung and breast cancers. Radiomics could constitute a non-invasive approach to unravel the origin of secondary brain lesions. This work aims to develop and validate a supervised machine learning model based on radiomic features extracted from brain MRI images to distinguish between metastases originating from breast and non-small cell lung cancers (NSCLC).

Materials/Methods: We retrospectively analyzed T1 Gadolinium-enhanced MR datasets for 79 patients with brain metastases from NSCLC and breast primaries, acquired from January 2013 to December 2018. A total of 98 lesions were identified. Tumors were manually contoured and 49 radiomic features were extracted for each brain lesion. The radiomics dataset was loaded to Python and was divided into training (70%) and validation (30%) sets. Three feature selection algorithms were used: Variance Threshold, Linear SVC, and Extra Trees Classifier. Six classification algorithms were compared for accuracy (ACC): Logistic Regression, Linear Discriminant Analysis, K-Nearest Neighbors, Classification and Regression Trees, Gaussian Naive Bayes, and Support Vector Machines. Ten-fold cross validation was done. The best performing algorithm was used to build the fitting model and was tested on the validation set for accuracy, F1 score, and area under the receiver-operating characteristic curve (AUC).

Results: The best model performance derived from combining Extra Trees Classifier and K-Nearest Neighbors algorithms (ACC = 0.68; Std. dev = 0.1). Eighteen relevant features were selected. The model was applied to the validation set and achieved a prediction accuracy of 0.83, an F1 score of 0.67 for breast and 0.89 for NSCLC, and an AUC of 0.81.

Conclusion: We developed an imaging-based non-invasive method using a machine learning approach to differentiate brain metastases secondary to NSCLC from those arising from breast cancers. Our study suggests that radiomic quantitative image analysis in this setting could provide useful site-specific correlates that could be validated in larger multicohort studies.

Author Disclosure: T.H. Felefly: None. S. Achkar: None. T. Lteif: None. J. Barouky: None. D. Nehme Nasr: None.

Samir Achkar, MD

Gustave Roussy Institute

Presentation(s):

Send Email for Samir Achkar


Assets

MO_19_2320 - A Machine Learning Model using MRI-based Radiomic Features to Predict Primary Site for Brain Metastases



Attendees who have favorited this

Please enter your access key

The asset you are trying to access is locked. Please enter your access key to unlock.

Send Email for A Machine Learning Model using MRI-based Radiomic Features to Predict Primary Site for Brain Metastases