Premium
Characterization of Adrenal Lesions on Unenhanced MRI Using Texture Analysis: A Machine‐Learning Approach
Author(s) -
Romeo Valeria,
Maurea Simone,
Cuocolo Renato,
Petretta Mario,
Mainenti Pier Paolo,
Verde Francesco,
Coppola Milena,
Dell'Aversana Serena,
Brunetti Arturo
Publication year - 2018
Publication title -
journal of magnetic resonance imaging
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.563
H-Index - 160
eISSN - 1522-2586
pISSN - 1053-1807
DOI - 10.1002/jmri.25954
Subject(s) - c4.5 algorithm , artificial intelligence , mcnemar's test , medicine , feature selection , machine learning , population , computer science , naive bayes classifier , radiology , nuclear medicine , support vector machine , mathematics , statistics , environmental health
Background Adrenal adenomas (AA) are the most common benign adrenal lesions, often characterized based on intralesional fat content as either lipid‐rich (LRA) or lipid‐poor (LPA). The differentiation of AA, particularly LPA, from nonadenoma adrenal lesions (NAL) may be challenging. Texture analysis (TA) can extract quantitative parameters from MR images. Machine learning is a technique for recognizing patterns that can be applied to medical images by identifying the best combination of TA features to create a predictive model for the diagnosis of interest. Purpose/Hypothesis To assess the diagnostic efficacy of TA‐derived parameters extracted from MR images in characterizing LRA, LPA, and NAL using a machine‐learning approach. Study Type Retrospective, observational study. Population/Subjects/Phantom/Specimen/Animal Model Sixty MR examinations, including 20 LRA, 20 LPA, and 20 NAL. Field Strength/Sequence Unenhanced T 1 ‐weighted in‐phase (IP) and out‐of‐phase (OP) as well as T 2 ‐weighted (T 2 ‐w) MR images acquired at 3T. Assessment Adrenal lesions were manually segmented, placing a spherical volume of interest on IP, OP, and T 2 ‐w images. Different selection methods were trained and tested using the J48 machine‐learning classifiers. Statistical Tests The feature selection method that obtained the highest diagnostic performance using the J48 classifier was identified; the diagnostic performance was also compared with that of a senior radiologist by means of McNemar's test. Results A total of 138 TA‐derived features were extracted; among these, four features were selected, extracted from the IP (Short_Run_High_Gray_Level_Emphasis), OP (Mean_Intensity and Maximum_3D_Diameter), and T 2 ‐w (Standard_Deviation) images; the J48 classifier obtained a diagnostic accuracy of 80%. The expert radiologist obtained a diagnostic accuracy of 73%. McNemar's test did not show significant differences in terms of diagnostic performance between the J48 classifier and the expert radiologist. Data Conclusion Machine learning conducted on MR TA‐derived features is a potential tool to characterize adrenal lesions. Level of Evidence : 4 Technical Efficacy : Stage 2 J. Magn. Reson. Imaging 2018.