A Comparative Analysis of Lung Image Classification using different Classification Techniques

Abstract

The urgent need for precise and prompt detection is further underscored by the fact that lung cancer is still the largest killer in the cancer category. In order to identify and categorize lung scans as normal or abnormal, this study suggests three separate image categorization techniques. Before anything else, lung images are preprocessed with Histogram Equalization to make them clearer and noise-free. After that, we use feature extraction methods, and then we use RF with Generalized Discriminant Analysis (GDA) to pick the best subset of features. During the classification stage, four classifiers are used: K-Nearest Neighbour (KNN), Naïve Bayes (NB), Neural Network (NN), and Random Forest (RF). Based on the comparison, the RF-GDA approach achieves better classification accuracy than the state-of-the-art methods. Second, computed tomography (CT) scans can be used to find early stages of lung cancer. Before feature extraction, the CT images are preprocessed using isotropic diffusion. The HOG approach is then used. One way to train a Neuro Fuzzy Classifier with Binary Cuckoo Search (NFCBCS) to better detect cancerous growth is to employ adjusted weights. The results of this study provide credence to the idea that image processing methods could be useful in the fight against lung cancer. The third tactic is the DF-PTDNN CAD model, which stands for Deep Features with Parameter-Tuned Deep Neural Networks. Prior to feature extraction using DenseNet121 and Local Binary Patterns (LBP), the model undergoes contrast augmentation and pre-processing based on Gaussian filtering. Various forms of lung cancer are detected and categorized by a deep neural network that employs a soft max classifier. Hyperparameters are fine-tuned using the quasi-oppositional moth swarm optimization (QOMSO) method, which yields remarkable results: sensitivity of 98.81%, specificity of 97.41%, and accuracy of 99.85%.