Colorectal polyps are precursors to colorectal cancer, making their early detection and classification crucial for effective treatment. This study focuses on developing Convolutional Neural Networks (CNNs) to classify the shapes of colorectal polyps in endoscopic images. Various CNN architectures, including ResNet101, VGG16, and a custom CNN, were trained and evaluated on the Kvasir dataset, which consists of annotated endoscopic images classified into flat, sessile, and pedunculated polyps. The dataset was divided into 75% for training and 25% for testing, with preprocessing techniques such as resizing, normalization, and augmentation applied to enhance model performance. Transfer learning was leveraged to enhance feature extraction, and the models were evaluated using metrics such as accuracy, recall, precision, and F1-score. Experimental results showed that ResNet101 achieved the highest accuracy of 95.15%, followed by VGG16 (94.85%) and the custom CNN (94.24%). Models such as InceptionV3 and Xception exhibited comparatively lower performance. These findings suggest that CNNs, particularly ResNet101, can effectively classify colorectal polyp shapes, aiding in early diagnosis and treatment. Future work will focus on hyperparameter optimization, ensemble learning, and advanced deep learning techniques to further enhance classification accuracy.