Deep Learning-Based Lung Cancer Detection and Classification with Hybrid Sampling for Imbalanced Data

Cancer is a disease that invades and destroys healthy tissue and disrupts vital bodily functions, often leading to high mortality rates. Detection and classification of lung cancer is a challenging task, and improving the recovery and survival rates of patients with lung cancer is crucial. Existing research focuses on computer-assisted diagnosis for effective detection and classification outcomes but fails to address imbalanced data distribution issues in the classification phase. This research tackles this issue and aims to provide a solution by using deep learning algorithms to mitigate bias towards the majority class, poor generalization, underfitting, overfitting, misleading performance metrics, and a high rate of false negatives. Images requiring preprocessing are sourced from LUNA60 and LIDC-IDRI. Preprocessing steps include image conversion to grayscale, resizing, normalization, and segmentation. For segmentation, a U-Net model is deployed to produce high-quality segmented images. The ResNet50 model is used for accurate classification. The misclassification rate is calculated using values derived from true positives, true negatives, false positives, and false negatives. If the misclassification rate is greater than zero, hybrid sampling is performed to balance the data using SMOTE for oversampling and CNN for undersampling. This process is repeated until the misclassification rate becomes zero. Finally, evaluation metrics such as precision, recall, F1 score, and AUC are used to optimize accuracy, which will aid in the detection and classification of lung cancer to ensure positive treatment outcomes for patients.