A Systematic Review of Deep Learning Architectures for Lung Cancer Detection and Classification

Lung cancer is the leading cause of cancer-related mortality worldwide, responsible for approximately 18% of all cancer deaths globally. The absence of early clinical symptoms significantly delays diagnosis, making timely and accurate automated detection systems a critical public health necessity. The rapid growth of deep learning technologies has opened transformative opportunities for automated lung cancer detection and classification from medical imaging data, particularly CT scans. However, despite impressive benchmark results, numerous methodological challenges and research gaps persist that hinder the transition of these models from laboratory settings to clinical practice. This paper presents a comprehensive systematic review of six representative deep learning-based studies for lung cancer detection and classification, published between 2018 and 2025. The reviewed works include four original research papers — spanning hybrid CNN-SVM models, transfer learning frameworks, CNN-LSTM architectures, and transformer-based segmentation systems — alongside two systematic review and survey papers that collectively examine over 30 models from the literature. A structured comparative analysis is conducted across eight key dimensions: paper type, datasets used, model architectures, classification task complexity, imaging modality, best reported accuracy, and other quantitative performance metrics including AUC, sensitivity, specificity, F1-score, and Dice coefficient. The critical analysis of each paper reveals recurring limitations across the field, including over-reliance on single benchmark datasets, restriction to binary classification tasks, absence of integrated segmentation and multi-class classification pipelines, minimal adoption of transformer-based models for classification, near-complete lack of model explainability mechanisms, and a universal absence of prospective clinical validation. Extending this analysis across all six reviewed papers, nine significant and cross-validated research gaps are systematically identified and documented. These gaps collectively define the design requirements for a next-generation lung cancer detection model: one that integrates 3D segmentation with multi-class subtype classification, employs transformer-based attention mechanisms, incorporates model interpretability, supports Low-Dose CT inputs, and is validated across diverse datasets and clinical environments. The findings of this review establish a rigorous evidence-based foundation for the development of a novel deep learning and image processing model for lung cancer detection, contributing to the ongoing effort to bridge the gap between computational performance and real-world clinical applicability.