This paper introduces DogHeart, a dataset comprising 1400 training, 200 validation, and 400 test images categorized as small, normal, and large based on VHS score. A custom CNN model is developed, featuring a straightforward architecture with 4 convolutional layers and 4 fully connected layers. Despite the absence of data augmentation, the model achieves a 72% accuracy in classifying cardiomegaly severity. The study contributes to automated assessment of cardiac conditions in dogs, highlighting the potential for early detection and intervention in veterinary care.