Cyberbullying is a serious online threat, especially in the form of image-based harassment on social media, with severe psychological impacts for victims. This systematic review critically discusses deep learning methods for image-based cyberbullying detection, providing a systematic evaluation of computational methods that tackle the multifaceted challenges of online harassment. The study examines various detection mechanisms, namely convolutional neural networks (CNNs), hybrid architectures, ensemble learning algorithms, and transfer learning methods, each having different mechanisms for detecting dangerous visual content. The rise of social media has spread the scope for digital abuse infinitely, and more sophisticated detection measures are becoming a necessity. Chief challenges in the detection of cyberbullying are posed by subjective interpretation of abusive content, insufficient high-quality datasets for training, and the swift evolution of harassment strategies online. Artificial intelligence appears to be the key tool here, with it showing immense promise in quickly scanning visual content for subtle patterns of online abuse. State-of-the-art architectures such as VGG16, ResNet50, and InceptionV3 demonstrate good potential in understanding intricate visual content and identifying possibly dangerous content. The research aggregates existing research to demonstrate remarkable development in multimodal detection approaches combining textual and visual analysis. Ultimately, the research highlights the pressing need for more advanced, context-aware algorithms and interdisciplinary collaboration to create safer digital spaces and limit the deep psychological effect of cyberbullying.