A Hybrid PSO–GA–SA Optimized CNN–KNN Framework for Automated Multiclass Lung Cancer Classification from CT Images

Abstract

Lung cancer is still one of the biggest causes of cancer deaths globally and it is essential to diagnose it accurately and timely from computed tomography (CT) images to enhance patient survival. While convolutional neural networks have been found to have good potential in automated CT-based classification, the performance of the networks is very much dependent on hyperparameter tuning, and the conventional hyperparameter tuning methods, such as manual search, grid search, and random search are computationally expensive and inefficient in high-dimensional space. In this study, a hybrid Particle Swarm Optimization–Genetic Algorithm–Simulated Annealing (PSO–GA–SA) methodology has been proposed to optimize CNN hyperparameters and CNN–KNN was designed to perform the automated multiclass lung cancer classification. The comprehensive preprocessing of the CT images is conducted to enhance the discriminative feature representation, such as resizing, grayscale, Gaussian filtering, Otsu thresholding, Canny edge detection, Sobel edge enhancement and channel fusion. The hybrid is a combination of PSO's global exploration capabilities, GA's evolutionary refinement capabilities and SA's local fine-tuning capabilities to achieve a balance between exploration and exploitation in optimization processes.
These optimized parameters were tested on CNN, MobileNetV2, DenseNet121, InceptionV3 and Hybrid CNN-KNN using Augmented IQ-OTHNCCD where accuracy, precision, recall, F1-score, confusion matrix and learning curve were used to assess their performance. The hybrid PSO–GA–SA optimized CNN–KNN model exhibited the best performance with 100% accuracy, 99% precision, 100% recall, and 98% F1-score, outperforming the individually optimized and stand-alone models, indicating that it has promising potential as a reliable clinical decision-support tool.