Secure and efficient internet of medical things framework for lung cancer diagnosis using transfer learning

The early diagnosis of Lung Cancer (LC) is crucial for improving the Survival Rate (SR) of patients. However, traditional Internet of Medical Things (IoMT) approaches face challenges regarding data privacy, storage, and security limitations. This research proposes a novel, secure, and efficient IoMT-based LC identification system. The hypothesis is that integrating effective encryption techniques, optimization, and Deep Learning (DL) will enhance the accuracy and efficiency of Lung Cancer Diagnosis (LCD) while ensuring data security. The methodology involves acquiring Computed Tomography (CT) scans through IoMT sensors. Similarly, the collected data is embedded using Odd Exponential Even Entropy-Lifting Wavelet Transform (OEEE-LWT)-based watermark embedding since it ensures data integrity, authenticity, and protection against tampering. Then, the embedded data are stored in the temporary storage layer. Afterward, the data is optimized by using the Transfer Function-Red Panda Optimization (TF-RPO) technique and then transferred to the gateway layer. Similarly, the data is encrypted using the Tangent Hyperbolic Chaotic Cryptosystem (THCC), ensuring confidentiality by protecting sensitive patient information from unauthorized access. Further, the encrypted data are transferred to the Lung Disease Diagnosis Model (LDDM). In LDDM, the Lung Cancer Prediction (LCP) is done by using the Transfer Learning-Saturated Wave-Convolutional Neural Network (TL-SW-CNN). Additionally, the Local Linear System-Fuzzy Inference System (LLS-FIS) predicts a 5-year SR based on extracted features like lung region and lymph nodules. Data balancing is performed using the Ranked Reverse Synthetic Minority Over-Sampling Technique (RR-SMOTE), and encryption is achieved with a Tangent Hyperbolic Chaotic Cryptosystem (THCC) to ensure security in distributed environments. Experimental results show that the proposed system achieved 98.01% classification accuracy and 97.93% precision, thus significantly reducing computational complexity. Also, when compared to the conventional models, such as CNN, RNN, DLNN, and ANN, the proposed TL-SW-CNN takes a minimum training time of 52389 ms, thereby reducing the computational time by 36.4%. The key contributions include enhancing data privacy, class balancing, LC prediction, and SR identification. Despite the strengths of the proposed system, attack detection has not been addressed, and this will be included in future work.