Deep learning-based approach to diagnose lung cancer using CT-scan images

Abstract

The work in this research focuses on the automatic classification and prediction of lung cancer using computed tomography (CT) scans, employing Deep Learning (DL) strategies, specifically Enhanced Convolutional Neural Networks (CNNs), to enable rapid and accurate image analysis. This research designed and developed pre-trained models, including ConvNeXtSmall, VGG16, ResNet50, InceptionV3, and EfficientNetB0, to classify lung cancer. The dataset was divided into four classes, consisting of 338 images of adenocarcinoma, 187 images of large cell carcinoma, 260 images of squamous cell carcinoma, and 215 normal images. Notably, The Enhanced CNN model achieved an unprecedented testing accuracy of 100%, outperforming all other models, which included ConvNeXt at 87%, VGG16 at 99%, ResNet50 at 94.5%, InceptionV3 at 76.9%, and EfficientNetB0 at 97.9%. The study of this research is considered the first one that hits 100% testing accuracy with an Enhanced CNN, demonstrating significant advancements in lung cancer detection through the application of sophisticated image enhancement techniques and innovative model architectures. This highlights the potential of Enhanced CNN models in transforming lung cancer diagnostics and emphasizes the importance of integrating advanced image processing techniques into clinical practice. • •The study proposes a deep learning-based approach for diagnosing lung cancer using CT-scan images, focusing on early detection and classification. • • Image enhancement techniques (e.g., HE, CLAHE, noise reduction) and preprocessing methods (e.g., augmentation) are applied to improve image quality and model robustness. • • ConvNeXt and other deep learning algorithms are used to train and analyze a large dataset from Kaggle, aiming to improve prediction accuracy and efficiency. • •The approach seeks to overcome limitations in manual diagnosis, providing faster, more reliable detection to aid radiologists in early lung cancer identification. • •The study addresses challenges in medical imaging interpretation, aiming to enhance survival rates by enabling timely and accurate lung cancer detection. • • Dataset Utilization : The study leverages a high-quality, annotated "Chest CT-Scan Images Dataset" from Kaggle, comprising 1,000 images labeled across four classes of lung cancer and normal cells, facilitating rigorous model training and testing. • • Image Enhancement : Image quality improvements use advanced techniques, such as median filtering for noise reduction, histogram equalization (HE), contrast-limited adaptive histogram equalization (CLAHE), and morphological operations, ensuring enhanced visibility and structure clarity in medical images for precise model predictions. • • Data Augmentation : Resizing, shearing, flipping, and rescaling techniques expand training data diversity and bolster model robustness, critical for improved generalization on unseen lung cancer data. • • Enhanced CNN Architecture : A custom CNN model with eight layers and dropout for overfitting control was developed, alongside pre-trained models (ConvNeXtSmall, ResNet50, VGG16, EfficientNetB0, and InceptionV3) for effective multi-class classification. • • Pre-trained Model Integration : Optimized pre-trained models were enhanced with dropout, pooling, and fully connected layers, trained and validated on distinct data splits, and evaluated for accuracy, loss, recall, and precision to ensure optimal lung cancer classification. • • Experimental Outcomes : Two experiment sets evaluated CNNs from scratch and pre-trained models, executed on Python in Kaggle’s environment, delivering competitive and reliable lung cancer detection metrics.