Enhanced Lung Cancer Detection Using Combined Edge–Threshold Segmentation And Fine-Tuned CNN Model

Abstract

The use of computer tomography (CT) imaging has become commonplace since it provides cross-sectional images of internal structures in the diagnosis of lung disease. It is more sensitive than traditional radiography, but even a single CT-use still subjects one to the risk of misdiagnosis, because artefacts and low contrast can hide nodules in their early stages. In the recent literature, it is emphasized that automated tools are required to complement the images and isolate the lung area and categorize nodules as benign or malignant. The use of otsu and adaptive thresholding has stayed in use to isolate the foreground and the background but thresholding in most cases does not work well with small lesions or when the background is not uniform. DenseNet and ResNet machine learning models are capable of producing good classification, but they need fine-grained segmentation and significant datasets. As a result, there exists a gap in techniques combining edge-based and threshold-based segmentation with a hybrid classifier. The proposed work presents a combination of a computer-aided diagnostic pipeline of lung cancer. The first stage is pre-processing where noise is removed, morphology, and adaptive histogram equalization are performed and then a merged segmentation is performed where an edge detector is used to highlight the nodule boundaries and an adaptive threshold is used to segregate lung tissue. Features are extracted using various Convolutional Neural Network (CNN) architectures and classified through a fine-tuned CNN-based classification framework with optimized hyper-parameter settings. The proposed hybrid model incorporates hyper-parameter optimization techniques including learning rate tuning, batch size optimization, dropout regularization, Adam optimizer configuration, and epoch selection to enhance classification performance and model generalization. The hybrid method was tested using a benchmark dataset, where the accuracy and F1-score achieved 99.24%, outperforming individual CNN models. The improvements were made due to the joint segmentation approach and the use of fine-tuned CNN architecture for final classification. The results suggest that the use of edge and threshold information together with a fine-tuned CNN model would significantly enhance early lung cancer detection.