Enhanced Lung Cancer Detection and Classification with mRMR-Based Hybrid Deep Learning Model

Abstract

Among the most common and fatal malignant tumors worldwide is lung cancer (LC). Generally, it has a poorer five-year survival rate than many other well-known tumors. Pulmonary nodules in the lung are an indicator of the most deadly and lethal kind of lung cancer. Improving a patient's chances of survival requires early detection and evaluation of lung cancer. In the field of lung cancer, broad use of deep machine learning techniques has led to notable progress in recent years in reaching high performance in early diagnosis and prognostic prediction. This study presented a novel hybrid deep learning model by applying the method, of two pre-trained deep model architectures ResNet101 and SqueezeNet obtain feature mappings from the CT images in the dataset. Out of the seven deep-learning CNN architectures that were tested, these two were chosen for evaluation based on their superior performance. Minimizing Redundancy and Maximize Relevance (mRMR) is also used to extract the best features from both of models in order to improve the computational efficiency and performance of the proposed technique. As a result, characteristics that have little bearing on accuracy are removed. All features are ranked to create a new set of feature maps. Next, the technique of feature concatenation is implemented. The best feature map is obtained and then classified using SVM and KNN, two machine learning (ML) classifiers.  The accuracy of the newly presented hybrid model was 99.09% with SVM. The results of the experiment show that the suggested hybrid model performed exceptionally well in terms of accuracy on the IQ-OTH/NCCD dataset.