MSGD-ResNet18 combines a modified stochastic gradient descent optimizer with ResNet18 architecture for CIFAR-10 image classification. The pipeline flows from 3×32×32 input through convolutional feature extraction, batch normalization, and ReLU activation, then through four residual layers (64, 128, 256, 512 channels) with integrated MSGD optimization at each stage. Global average pooling reduces spatial dimensions before a fully connected layer maps 512 features to 10 CIFAR-10 classes. This architecture demonstrates how custom optimizers can be embedded within residual blocks to improve gradient flow and training stability on small-scale image datasets. Fork and customize this diagram to experiment with different optimizer placements, channel configurations, or alternative pooling strategies for your own CNN designs.