MRI Radiomics of Breast Cancer: Machine Learning-Based Prediction of Lymphovascular Invasion Status

• Published in: Academic radiology Vol. 29 Suppl 1; p. S126
• Main Authors: Kayadibi, Yasemin, Kocak, Burak, Ucar, Nese, Akan, Yesim Namdar, Yildirim, Emine, Bektas, Sibel
• Format: Journal Article
• Language: English
• Published: United States 01.01.2022
• Subjects: Breast Neoplasms - diagnostic imaging; Breast Neoplasms - pathology; Female; Humans; Lymph Nodes - diagnostic imaging; Lymph Nodes - pathology; Machine Learning; Magnetic Resonance Imaging - methods; Neoplasm Invasiveness - diagnostic imaging; Reproducibility of Results; Retrospective Studies; Texture analysis; Breast cancer; Lymphovascular invasion; Machine learning; Radiomics
• ISSN: 1878-4046
• DOI: 10.1016/j.acra.2021.10.026

In patients with breast cancer (BC), lymphovascular invasion (LVI) status is considered an important prognostic factor. We aimed to develop machine learning (ML)-based radiomics models for the prediction of LVI status in patients with BC, using preoperative MRI images. This retrospective study included patients with BC with known LVI status and preoperative MRI. The dataset was split into training and unseen testing sets by stratified sampling with a 2:1 ratio. 2D and 3D radiomic features were extracted from contrast-enhanced T1 weighted images (C+T1W) and apparent diffusion coefficient (ADC) maps. The reliability of the features was assessed with two radiologists' segmentation data. Dimension reduction was done with reliability analysis, multi-collinearity analysis, removal of low-variance features, and feature selection. ML models were created with base, tuned, and boosted random forest algorithms. A total of 128 lesions (LVI-positive, 76; LVI-negative, 52) were included. The best model performance was achieved with tunning and boosting model based on 3D ADC maps and selected four radiomic features. The area under the curve and accuracy were 0.726 and 63.5% in the training data, 0.732 and 76.7% in the test data, respectively. The overall sensitivity and positive predictive values were 68% and 69.6% in the training data, 84.6% and 78.6% in the test data, respectively. ML and radiomics based on 3D segmentation of ADC maps can be used to predict LVI status in BC, with satisfying performance.