A lightweight ALO optimized and learnable skip-connection integrated ResNet architecture for breast cancer diagnosis

Hiren Mewada, Ivan Miguel Pires, Hiren Kumar Thakkar, Amit Patel

Breast cancer is a global health concern, and early detection through screening programs is crucial for reducing mortality. Deep Convolutional Neural Networks (CNNs) are widely adopted for image classification, but their accuracy depends on the number of layers, structure parameters, and hyperparameters. This paper presents an optimized ResNet architecture for breast cancer classification from histopathological images, introducing learnable convolutional skip-connections to enhance feature fusion efficiency and reduce computational demands. The model architecture is optimized using Ant Lion Optimization (ALO) to automatically tune critical parameters, achieving an efficient 74-layer structure with just 5.2 million parameters. When evaluated on the BreaKHis dataset, the approach demonstrated superior performance compared to existing methods. For binary classification, the model achieved 96.30% accuracy, outperforming recent studies that used residual networks. In multi-class classification, the proposed model achieved an accuracy of 91.37%, surpassing conventional ResNet architectures while maintaining lower computational complexity. These results highlight the potential of the method to improve diagnostic accuracy in clinical settings, offering a balanced solution to the accuracy–efficiency trade-off in deep learning-based cancer detection.