AI-Driven Diagnostic Imaging: Hybrid CNN-GNN Models for Early Detection of Cancer from Pathological Images

Abstract: The early and accurate detection of cancer from histopathological images is crucial for the improvement patient outcomes in precision oncology because conventional diagnostic methods usually suffer from subjectivity and high variability, while traditional deep learning approaches, though effective, are limited in capturing both local morphological details and global tissue context simultaneously. In order to address this challenge, this study proposes a hybrid Convolutional Neural Network–Graph Neural Network (CNN–GNN) framework that integrates patch-level visual feature extraction with graph-based relational learning for cancer detection. The study adhered to the Agile approach and publicly available datasets, CAMELYON16 and CAMELYON17, were used, which consist of Whole-Slide Images (WSIs) and professional annotations of normal and metastatic tissue areas. Stain normalization, patch extraction, data augmentation, and graph construction were used as preprocessing steps, which provided both CNN and GNN pipelines with high-quality inputs. DenseNet121 was used in place of CNN backbone to extract patch embedding whereas Graph Convolutional Network (GCN) was used to learn the spatial and contextual relationship among patches. The last distinction came by combining CNN and GNN embedding by a multilayer perceptron classifier. The effectiveness of the given architecture was proven by experiment results. CNN model reached an accuracy of 88.9% with an F1-score of 89.2% and GNN model reached a higher accuracy of 90.7% and F1-score of 91.0%. The hybrid CNNGNN model notably outdid the two baselines, achieving a test accuracy of 95.4%, precision of 94.7%, recall of 95.9%, F1-score of 95.3% and AUC of 96.4%. Therefore, the hybrid CNNGNN model that is suggested provides a scalable, trustworthy, and clinically feasible solution to computational pathology. Along with attention mechanisms, enhanced GNN variants, and data on multiple institutions, future extensions could help to expand the overall generalizability and clinical uptake.