Experimental Insights into Multi-Class Lung Disease Detection from CT scans

Abstract

Lung diseases including cancer, pneumonia, and interstitial lung disorders still account for the largest proportion of mortalities globally, and early and accurate detection could significantly enhance patient care. CT imaging is the best option for detecting preliminary pulmonary abnormalities, but the time taken for manual interpretations is considerable and the outcomes vary from one clinician to another. This paper outlines an experimental review of the detection of multi-class lung diseases from CT scans involving the comparisons of classical machine learning, end-to-end deep learning, hybrid deep-feature classifiers, and ensemble techniques. The Data Preparation step involves the pipeline of data preprocessing, contrast enhancement using CLAHE, normalization, and augmentation for the CT images to be ready for analysis. Radiomic descriptors and convolutional neural networks of ResNet-50, DenseNet-121, and MobileNetV2, as well as other hybrid systems that integrate deep learning and classical systems for feature extraction. Assessments are made using precision, recall, F1-score, balanced accuracy, ROC-AUC, and Grad-CAM for interpretability. The analysis noted classical approaches were significantly outperformed by deep learning, with strong classification performance across Normal, Benign, and Malignant classes in ResNet-50 and DenseNet-121. Ensemble methods added to the robustness of the models achieving a macro-F1 of about 91% and an AUC of 0.96. Benign lesion detection is the most problematic because it shares overlapping features with malignant lesions. This review emphasizes the potential of attention-guided and hybrid methods in reconciling accuracy with interpretability, while also addressing fundamental concerns such as dataset heterogeneity, generalizability, and the need for clinical validation.