Histopathological image analysis is essential for cancer diagnosis, but accurately segmenting and classifying cancer cells is a complex task due to structural variations in medical images. This study implements a two-stage pipeline that integrates self-supervised learning with graph-based classification to enhance performance while minimizing reliance on labeled data. In the first stage, SimCLR—a contrastive learning framework—is integrated with EfficientNet-B4 to extract meaningful and invariant feature representations from unlabeled histopathological images. The extracted features are then used to construct a k-nearest neighbor graph, which serves as input to a multi-layer graph Attention Network (GAT) for classification. In the second stage, the graph is processed by a multi-layer Graph Attention Network (GAT), which dynamically weighs neighboring nodes to perform accurate classification. The architecture incorporates adaptive graph generation, efficient augmentation pipelines, and a cosine-annealed learning rate for training stability. The self-supervised nature of the method reduces the dependency on labeled data, offering a scalable and robust solution for histopathological analysis.