Deep learning based similarity-consistency abnormality detection (SCAD) model for classification of MRI patterns of multiple myeloma (MM) infiltration

With IRB approval, a total of 132 sagittal views of T1-weighted (T1W) sequence in the spinal MRI scans was collected from 67 patients at our institution and used in this study. We developed a similarity-consistency abnormality detection model (SCAD) for classification of MRI patterns that are associated with low and high risk of multiple myeloma (MM) disease. Our SCAD model consisted of five CNN structures: a generator, an encoder, and three discriminators. The generator was used to capture the distribution of training samples by mapping the given distributions to the distribution of training samples. The encoder mapped the distribution of training samples back to the given distribution to speed up the inference time. The three discriminators were designed to force the generator and the encoder to meet a cycleconsistency constraint. The five components are trained together following the min-max game in GAN. The MRI patterns of each vertebra (normal, focal, variegated and diffused) were provided by an experienced radiologist as reference standard. We trained our SCAD model using the vertebras with the normal pattern, and deployed the trained SCAD model to the three non-normal patterns. The results showed that, our SCAD model achieved a test AUC of 0.71, 0.79 and 0.88 in identifying the focal, variegated and diffused patterns from the normal pattern, respectively.