This research assesses the utility and validity of using simulated radar reflectivity to detect potential changes in linear and nonlinear mesoscale convective system (MCS) occurrence in the Midwest United States between the early and late 21st century using convection-permitting climate simulation output. These data include a control run and a pseudo-global warming (PGW) run that is based on RCP 8.5. First, using a novel segmentation, classification, and tracking procedure, MCS tracks are extracted from observed and simulated radar reflectivity. Next, a comparison between observed and the control run MCS statistics is performed, which finds a negative summertime bias that agrees with previous work. Using a convolutional neural network to perform probabilistic predictions, the MCS dataset is further stratified into highly organized, quasi-linear convective systems (QLCSs)-which can include bow echoes, squall lines, and line echo wave patterns-and generally less-organized, non-QLCS events. The morphologically stratified data reveal that the negative MCS bias in this region is largely driven by too few QLCSs. Although comparisons between the control run and a PGW run suggest that all MCS events are less common in the future (including QLCS and non-QLCS events), these changes are not spatially significant, whereas the biases between the control run and observations are spatially significant. A discussion on the importance and challenges of simulating QLCSs in convection-permitting climate model runs is provided. Finally, potential avenues of exploration are suggested related to the aforementioned issues. K E Y W O R D S climate modeling, mesoscale convective systems, convolutional neural network