Devising a comprehensive approach to diagnosing breast cancer subtypes automatically based on deep neural networks

This study investigates the process of analyzing immunohistochemical images of breast cancer. The study has contributed to solving the task of a standardized and objective approach to the quantitative assessment of immunohistochemical biomarkers, which would minimize inter-individual variability in assessments and could be computationally efficient for the analysis of biomedical images. This paper aims to balance model complexity and generalization by using evolutionary algorithms to tune deep neural networks for biomedical tasks, analyzing how network structure affects performance. Experiments were conducted on the segmentation of immunohistochemical images on 13 different architectures of neural networks. The evaluation was performed using five accuracy metrics, which allowed for an objective comparison of model performance. The use of a genetic algorithm to optimize the neural network architecture made it possible to adaptively find combinations of parameters, in particular the number of layers and the size of the base filter. The evolutionary approach enabled effective exploration of configuration space, which led to an increase in the Dice metric to 0.74. The resulting increase in accuracy indicates the model’s improved ability to segment images with different characteristics, demonstrating the practical effectiveness of the proposed approach for biomedical diagnosis tasks. The optimized architecture was used to design a system for diagnosing breast cancer automatically based on neural networks, in particular for the method of automatic diagnosis of breast cancer subtypes. That contributed to improving the accuracy of biomedical image analysis, which could help improve the diagnostic process in clinical practice