In response to the challenges of limited annotated lung cancer pathology images with complex cell morphology, we propose a lung cancer pathology image classification method based on contrastive learning of cell morphology features. This approach integrates highly reliable unlabelled data into the training data through contrastive learning to address the issue of insufficient labelled data. Building upon the foundation of k-nearest neighbour contrastive learning, we introduce farthest and nearest neighbour contrastive learning, simultaneously using the farthest and nearest neighbour images for contrastive learning. This enhances the performance of contrastive learning by increasing the difficulty of positive sample learning and diversifying the dataset. We employ a ResNet50 encoder with deformable convolutions and dynamic convolutions to improve the extraction of cell morphology features. Experimental results demonstrate that, especially when labelled data is scarce, our classification method effectively leverages cell feature information from both labelled and unlabelled cancer pathology images, resulting in superior classification performance compared to existing methods. When applied to 80%, 60%, and 40% labeled data, respectively, CCCM outperforms other approaches with superior performance metrics, reaching F1 Score of 0.9584, 0.9488, and 0.9386.