This chapter presents a comparative analysis of the lightweight Deep learning models DenseNet-Lite, MiniResNet, and MobileNet-Lite in the multiclass coronary stenosis grading problem. In general, the automatic classification of coronary stenosis has been addressed using binary classifiers, since the low number of available data. For multiclass classification, quantitative coronary stenosis grading has been categorized as no visible, mild, moderate and severe. The Deep Learning models were compared with well-known Machine Learning techniques in terms of accuracy, precision, recall and F1-score. The whole database contains 500 synthetic grayscale images per class, which were generated based on Bézier curves. To perform the training and testing of the methods, the database was partitioned into 80% training and 20% testing of grayscale images. According to the experimental results, the DenseNet-Lite achieves superior performance than the comparative methods obtaining an accuracy of 0.7925, precision of 0.7995, recall of 0.7924, and F1-score of 0.7945 using the test set of images. In terms of computational time, the DenseNet-Lite obtained an average time of 76 s in training and 0.0009 s per image in testing. In addition to the experimental results and based on the computational time, the DenseNet-Lite Deep Learning architecture can be a potential model to be part of systems that perform computer-aided diagnosis in cardiology.