A NOVEL MULTI-SCALE ATTENTION-BASED DEEP LEARNING ARCHITECTURE FOR AUTOMATED LUNG CANCER SCREENING, EARLY TUMOR DETECTION, AND INTELLIGENT CLINICAL DECISION SUPPORT USING HIGH-RESOLUTION COMPUTED TOMOGRAPHY IMAGE ANALYSIS

Abstract

Lung cancer remains one of the leading causes of cancer-related mortality worldwide, primarily because a large proportion of cases are diagnosed at advanced stages when treatment options are limited. Early identification of malignant pulmonary nodules from computed tomography (CT) images can substantially improve patient survival and clinical outcomes. This study proposes a novel multi-scale attention-based deep learning architecture for automated lung cancer screening, early tumor detection, and intelligent clinical decision support using high-resolution computed tomography image analysis. The proposed framework is designed to capture both local fine-grained features and global contextual information through multi-scale feature extraction and attention mechanisms, thereby enhancing the discrimination between benign and malignant pulmonary nodules. The experimental study was conducted using the publicly available LIDC-IDRI (Lung Image Database Consortium and Image Database Resource Initiative) and LUNA16 (LUng Nodule Analysis 2016 Challenge) datasets, comprising more than 1,180 thoracic CT scans and over 8,800 annotated pulmonary nodules reviewed by experienced thoracic radiologists. CT images were preprocessed through intensity normalization, lung parenchyma segmentation, noise reduction using Gaussian filtering, image resizing, and data augmentation techniques including rotation, flipping, scaling, and contrast enhancement to improve model robustness and reduce overfitting. The proposed architecture was implemented using Python 3.11, TensorFlow 2.16 and Matplotlib, while training and evaluation were performed on an NVIDIA GPU-enabled computing platform with five-fold cross-validation. Experimental results demonstrate that the proposed framework outperformed several state-of-the-art deep learning models, including ResNet50, DenseNet121, EfficientNet-B4, and Vision Transformer. The proposed model achieved an accuracy of 98.4%, precision of 98.1%, recall (sensitivity) of 98.3%, specificity of 97.9%, F1-score of 98.2%, ROC-AUC of 0.996, and a Dice Similarity Coefficient of 97.5% for pulmonary nodule localization. Statistical analysis confirmed the superiority of the proposed method (p < 0.001), while Grad-CAM visualization demonstrated accurate localization of suspicious tumor regions, improving the interpretability of automated predictions. The proposed multi-scale attention-based framework provides a reliable and interpretable computer-aided diagnostic solution that enhances screening efficiency, supports radiologists in early lung cancer diagnosis, reduces false-positive findings, and facilitates intelligent clinical decision-making. The developed framework has strong potential for integration into next-generation smart healthcare systems and hospital radiology workflows to improve diagnostic accuracy and patient management.