Improving Early Detection and Diagnosis of Lung Cancer Using an Enhanced Ensemble Deep Learning Model on CT Images

Abstract

Purpose: Lung cancer is the most common and deadliest type of cancer that is the cause of one million deaths around the world every year. Due to the present level of medical research, identifying lung tumors on chest Computed Tomography (CT) images has become a significant process in modern medicine. Enhancing treatment and reducing lung cancer mortality can be achieved by promptly identifying and accurately diagnosing suspected malignant lung tumors. While many deep learning algorithms have been developed recently for the classification of lung cancer, achieving high accuracy in lung cancer classification is still a challenge. An advanced deep learning technique is designed to boost the effectiveness of early lung cancer diagnosis.

Materials and Methods: In this research, we have proposed an enhanced ensemble deep-learning model for lung cancer classification and segmentation. Initially, we carried out an extensive pre-processing process, including image resizing, noise reduction, and contrast enhancement, to enhance the image quality. The problem of small sample size is addressed by applying conventional data augmentation techniques like flipping, rotating, zooming, and shearing. Next, seven statistical features are retrieved using Improved Empirical Wavelet Transforms (IEWT). After feature extraction, the Enhanced ResNeXt model is used to classify lung cancer into normal, malignant, and benign classes. The interested region of the lung tumor is segmented using the Modified ShuffleNetV2 model. Individuals are classified into normal, malignant, and benign categories based on the presence of lung cancer. The experiments are performed on the benchmark datasets LIDC-IDRI and IQ-OTH/NCCD.

Results: For lung cancer classification, the proposed Enhanced ResNeXt model achieves an exceptional model accuracy of 99.43% for the IQ-OTH/NCCD dataset and 99.37% for the LIDC-IDRI dataset. Furthermore, the proposed Modified ShuffleNetV2 model effectively segments lung tumor regions, achieving an Intersection over Union (IOU) of 98.43% and a Dice Similarity Coefficient (DSC) of 97.24% on the LIDC-IDRI dataset, and an IOU of 97.05% and DSC of 96.23% on the IQ-OTH/NCCD dataset, demonstrating its robustness and accuracy in delineating tumor boundaries. The expected outcomes show that the accuracy and efficiency of our proposed ensemble deep learning model outperform other CNNs.

Conclusion: The proposed models beat existing CNN-based models in terms of speed and number of training parameters, which means that using CT scan images to diagnose lung cancer automatically is a suitable option and a strong selection for extensive use in medical environments.