Cloud assisted Multimodal Lung Cancer Classification Using an explainable Deep Learning approach

Abstract

Lung cancer ranks as the most common cause of death among all cancer types globally. Conventionally, CT scans and MRI were utilized to detect the lung modules, but it is time consuming and prone to errors, so this proposed model developed a novel deep learning approach to classify the lung cancer, which utilizes a PET, genomic, and clinical images to classify the lung modules. Initially, missing values are balanced utilizing mean imputation, then the data is normalized using min-max scaling. In the pre-processed images, the features are extracted employing PyRadiomics, Principal Component Analysis (PCA), and t-distributed stochastic Neighbor Embedding (t-SNE), which is utilized to reduce the noise. Additionally, the deep features are extracted employing Residual AlexNet, and the clinical data is encoded using One-hot encoding. Additionally, the Explainable cross-module attention deep fusion network pruning transformer (CDF-NPT) framework is used to classify the output. Shapely Additive Explanation (SHAP) is the framework utilized for feature attribution. Finally, the data is encrypted using a Secured dual stage encryption system technique using ElGamal ECC to secure the patients records. Further, the encrypted data is stored in the cloud platform successfully. Performance of this proposed system is evaluated utilizing metrics such as accuracy, precision, recall, F1-score, Encryption time, Execution time, ROC curve, Training and testing accuracy, and so on.  The accuracy of this proposed model is 98.47% for TCGA and 99.01 % for the TCIA dataset, and the encryption time of 69 ms and 64 ms, respectively. The overall results indicate the proposed model outperforms the existing models.