Deep Learning to Automate Reference-Free Image Quality Assessment of Whole-Heart MR Images

• DOI: 10.1148/ryai.2020190123
• Publication Date: 2020-05-27T13:58:28Z
• Authors: Davide Piccini; Robin Demesmaeker; John Heerfordt; Jérôme Yerly; Lorenzo Di Sopra; Pier Giorgio Masci; Juerg Schwitter; Dimitri Van De Ville; Jonas Richiardi; Tobias Kober; Matthias Stuber
• Keywords: Quality Score; Concordance

To develop and characterize an algorithm that mimics human expert visual assessment to quantitatively determine the quality of three-dimensional (3D) whole-heart MR images.In this study, 3D cardiac MRI scans from 424 participants (average age, 57 years ± 18 [standard deviation]; 66.5% men) were used generate image algorithm. A deep convolutional neural network for (IQ-DCNN) was designed, trained, optimized, cross-validated on a clinical database 324 (training set) scans. On separate test set (100 scans), two hypotheses tested: (a) can assess in concordance with as assessed by human-machine correlation intra- interobserver agreement (b) IQ-DCNN may be monitor compressed sensing reconstruction process where progressively improves. Weighted κ values, disagreement counts, Krippendorff α reliability coefficients reported.Regression performance within range very good (R2 = 0.78, 0.67). The during correlated cost function at each iteration successfully applied rank results expert.The proposed trained mimic images. reading, capable automatically comparing different reconstructed volumes.Supplemental material is available article.© RSNA, 2020.