Enhancing CT-Based Lung Cancer Diagnosis with CNN and LSTM Architecture

New case of lung cancer is among the leading causes of cancer related deaths in the world and therefore early detection improves prognosis. In this paper, the author presents a new deep learning model, which combines CNN and LSTM for better lung cancer detection. The CNN component extracts the spatial features out of individual slices of the CT, while the LSTM extracts temporal dependencies over several consecutive slices to let the model take advantage of both spatial and temporal patterns. The proposed CNN-LSTM model can reach 92.1% accuracy on LIDC-IDRI dataset, it significantly outperforms the standalone CNN with 87.0% and LSTM with 85.3%. Indeed, the current model achieved high accuracy (90.2%), recall (91.5%), and F1- score (90.8%), which suggests that it is effective in minimizing the number of false negatives and false positives that are critical in clinical diagnosis. It has a high accuracy in distinguish malignant from benign nodule with the AUC of 0.93 as well. The ablative studies supported the joint operations of the CNN and LSTM, and the application of the integrated architectural design was found to show improved performance. Hyperparameter tuning, with the learning rate at 0.001 and the utilizing of 128 LSTM units, brought improvements to learn statistical models. Grad-CAM graphics enhanced interpretability as focal points on the regions associated with malignancy were enlarged. However, problems like high computation time and requirement of large amounts of labeled data still exist even in such a well-developed model. To the present study, this paper provides a theoretical possibility of early and accurate diagnosis of lung cancer to improve the application of AI in healthcare.