Deep Learning approach for breast cancer discovery Via gene correlation

Background Identifying breast cancer early is crucial for successful treatment and reducing the risk of developing the disease. Many people fail to detect the illness in its early stages, which increases the chances of progression and higher mortality. Breast cancer, common among women, can be effectively treated if caught early before it advances. However, current methods for early diagnosis have limitations. Objective This study aims to predict breast cancer early using a genetic dataset from The Cancer Genome Atlas (TCGA), which includes 1078 patients and 21 genes. Methodology A hybrid approach combining Convolutional Neural Networks (CNN) with Bidirectional Long Short-Term Memory (BiLSTM) was used. The Lasso method was applied to identify genes significantly affecting disease risk. The hybrid model was trained on a subset of data containing only the key genes. Bayesian optimization techniques were used to determine the best hyperparameters. To enhance the model's generalization and ensure consistent performance, 10-fold cross-validation was performed during training. Model performance was evaluated using accuracy, precision, recall, F1-score, and the confusion matrix. Results The hybrid model combining CNN and BiLSTM achieved the highest mean accuracy of 99.27%, while the CNN alone reached an accuracy of 98%. Sigmoid and ReLU were the activation functions used. Conclusion Incorporating specific genes and their interactions significantly improved prediction accuracy. Conclusion: The best performance was achieved by selecting the most influential genes and their interactions and tuning hyperparameters with the hybrid model. These key genes proved most effective in identifying breast cancer.