Volume - 7 | Issue - 3 | september 2025
Published
01 August, 2025
Lung cancer remains one of the leading causes of death worldwide, mainly due to delayed diagnosis and treatment. Lung nodules must be properly and quickly recognized as benign or malignant to improve survival. This work presents a deep learning approach for segmentation using the Residual U-Net (ResiU-Net) architecture improved with Multiscale Aggregation (MSA) focusing on accurately identifying lung nodules from CT scans. A total of 1,190 labeled CT scan slices from the publicly available IQ-OTHNCCD dataset, including the benign, normal, and malignant classes, are used to test the model. To enhance feature representation and data balance, advanced preprocessing methods such as data augmentation and standardization were used. The new ResiU-Net architecture makes use of residual connections to overcome vanishing gradients and achieve contextual depth, while the MSA block enables the collection of global and fine-grained information required for the identification of small nodules (as small as 3 mm). A rigorous 5-fold cross-validation technique was used to test the model's performance. The strength of segmentation was evaluated using metrics such as F1-Score, Accuracy, Dice Score, and Intersection over Union (IoU). The method demonstrated better detection accuracy and excellent generalization compared to baseline U-Net models with an F1-score of 0.8385 and a best IoU of 0.7971 and validation IoU of 0.7300. The model's ability to accurately recognize nodules of varying sizes and shapes has been established by the testing results. The goal of the project is to help design a clinical decision-support system that is accurate, automated, and affordable for monitoring and early-stage lung cancer diagnosis.
KeywordsPulmonary Nodule Segmentation Deep Learning U-Net Resnet-152 Computed Tomography (CT) Images Medical Image Processing Convolutional Neural Networks (CNNs)
