Breast cancer is among the most prevalent and lethal illnesses in females around the world, and its early diagnosis and accurate diagnosis are of extreme necessity for optimal treatment and improved survival rates. This study utilized a recent breast cancer dataset from Mendeley with 125 Cancer images and 620 Non-Cancer images to examine five pre-trained convolution neural network models; MobileNetV2, Xception, VGG19, DenseNet201, and ResNet50V2 and a proposed Hybrid model (DenseNet+VGG19) were trained and compared based on precision, recall, F1 score, and accuracy. Experimental results certified that all the models were good with 97% accuracy for MobileNetV2, 97.33% accuracy for Xception, 98.83% accuracy for VGG19, and 99% accuracy for DenseNet201 and ResNet50V2. The Hybrid model of this work achieved highest accuracy of 99.33% with near perfect precision and recall values for both cancer and non-cancer types, demonstrating improved feature extraction and classification capability. Misclassification analysis showed that the Hybrid model had the fewest errors (4 cases), followed by the other models. ROC analysis also indicated superb predictive performance, where all models attained an AUC of 1.00, suggesting nearly perfect discrimination among classes. Extrinsic testing with the INbreast database substantiated the robustness of the proposed Hybrid model, and it achieved a general accuracy of 89.61%, precision, recall, and F1 scores of 0.90, 0.82, and 0.86 for cancer class, respectively. The hybrid model designed here was extremely effective and stable in the auto-detection of breast cancer and has future potential for clinical applications and optimization using bigger datasets and light model optimization.