@article{Alrasheedi2025, 
author = {Masad A. Alrasheedi and Asamh Saleh M. Al Luhayb and Abdulmajeed A. R. Alharbi},
title = {Deep learning framework for early diagnosis of lung cancer using multi- modal medical imaging},
year = {2025},
journal = {AIMS Mathematics},
volume = {10},
number = {12},
pages = {29815-29852},
keywords = {lung cancer diagnosis, multi-modal medical imaging, cross-modal transformer, uncertainty calibration, prototype reasoning, concept bottleneck, deep learning},
url = {https://www.sciopen.com/article/10.3934/math.20251310},
doi = {10.3934/math.20251310},
abstract = {Early and accurate diagnosis of lung cancer remains challenging due to the heterogeneity of tumor morphology and the variability across imaging modalities. This study proposed a deep learning framework that integrated computed tomography (CT), positron emission tomography/computed tomography (PET/CT), and chest X-ray (CXR) within a unified multi-modal transformer architecture for early lung cancer detection. The framework employed modality-specific encoders combining convolutional and state-space blocks to extract spatial-frequency representations, followed by a gated cross-modal fusion transformer designed to align heterogeneous features and handle missing modalities through mixture-of-experts routing and low-rank imputation. Multi-task heads were jointly optimized for nodule detection, segmentation, malignancy classification, and survival risk prediction. Explainability was embedded through concept bottlenecks, prototype reasoning, gradient-based attribution, and counterfactual concept editing, offering case-level interpretability and clinically meaningful evidence maps. Uncertainty was estimated via Monte-Carlo dropout, deep ensembles, and temperature scaling to ensure calibrated confidence estimates and defer-to-expert safety decisions. Lung image database consortium and image database resource initiative (LIDC-IDRI) (CT), the cancer imaging archive (TCIA) (PET/CT), and national lung screening trial (NLST) (CXR) benchmark datasets revealed that our methods work better than the best methods available. The proposed technique yielded Dice scores of 0.879, 0.872, and 0.876, together with AUC values of 0.944, 0.952, and 0.938, and an expected calibration error (ECE) of 0.02 across all modalities. Under domain shift, cross-dataset analysis showed substantial generalization (   A  U  C  >  0.92). A generalizable framework for multi-modal diagnostics made it possible to use AI to help with lung cancer screening in a way that was clear, trustworthy, and scalable.}
}