Optimizing breast cancer classification with EGERIA and reinforcement learning

In recent years, breast cancer (BC) diagnosis has drawn attention from the biomedical informatics community in the healthcare industry. Manual evaluation of histopathology images is both time-consuming and resource-intensive. Training large deep neural networks (DNNs) is computationally demanding and requires significant computing power. To address this issue, we propose an advanced classification framework for the most precise classification of breast cancer. The backbone network, DenseNet121, is trained via transfer learning to extract deep features from breast cancer images. These features are further refined using Dynamic Graph Convolutional Filters (DGCNs) to capture spatial dependencies. To enhance this framework, Efficient Knowledge-Guided Layer Freezing (EGERIA) is incorporated, accelerating training by freezing converged layers. Soft Actor-Critic is employed to address misclassification issues by dynamically adjusting model parameters. We used the BreakHis dataset, which is publicly available, to conduct extensive experiments with the proposed architecture. Experimental results demonstrate that our method outperforms existing approaches, achieving an accuracy of 99.65 %. The proposed framework not only achieves high accuracy and efficiency in breast cancer classification but also ensures robust performance across numerous evaluation metrics, presenting a promising tool for medical practitioners.