Spatial-Sequential Analysis for Breast Cancer Diagnosis: Integrating CNN and LSTM Architectures

Diagnosis of breast cancer is a extremely critical task in medical imaging, for which accurate and efficient computational methods are required. In this thesis, a novel hybrid deep learning framework that integrates Convolutional Neural Networks (CNNs) to extract spatial features and Long Short-Term Memory (LSTM) networks to recognize sequential patterns improves breast cancer classification. Evaluation of the proposed model on benchmark mammography datasets shows that the model achieves an accuracy of 92.8 %, compared to standalone CNN (89.3 % and LSTM (85.2)). Precision wise, the hybrid approach received 92.3 to 88.7 and 84.9 for CNN and LSTM, respectively, and recall at 92.7, 89.1, and 85.5. Finally, the proposed hybrid model achieved an F1score of 92.5% that is the balanced tradeoff between precision and recall. In addition, Receiver Operating Characteristic (ROC) analysis was performed to validate the effectiveness of the model, which showed a high value of 0.96 for the AUC score for hybrid model compared with CNN (AUC 0.91) and LSTM (AUC 0.88) score that prove its ability to separate malignant and benign cases better than other models. The training and validation accuracy curves were plotted and we observed a steady improvement in both of them and training loss reduced from 0.60 to 0.10 and validation loss reduced from 0.65 to 0.15 showing that it is good at learning and generalization. Furthermore, these results prove that the CNN-LSTM hybrid model is a promising solution for automated breast cancer diagnosis through the ability to learn spatial and sequential features to help diagnose more accurately and robust manner. Future work involves improving interpretability through use of explainable AI techniques, as well as evaluation across larger datasets.