Background/Objectives: Lung cancer is one of the most common and lethal malignancies worldwide. Early detection remains challenging due to its variable biological behavior. Computed tomography (CT) is the primary imaging method used for early detection. However, the manual interpretation of CT scans is constrained by several challenges such as reliance on expert experience, increasing clinical workload, and considerable variability among observers. Methods: This study introduces an explainable transformer-based framework capable of distinguishing among the three principal clinical categories of lung cancer (small-cell lung cancer, non-small-cell lung cancer, and normal) while simultaneously generating automated radiology reports from CT images. In contrast to conventional single-slice methodologies, the proposed model employs a multi-slice volumetric encoding strategy that captures spatial continuity and anatomical relationships across the CT slices. Visual features extracted by a ViT-based encoder are transformed into a compact patient-level representation through a Learnable Query Attention Pooling (LQAP) mechanism, and this unified representation is subsequently used for both three-class prediction and report generation with a GPT-2-based decoder. To enhance explainability, slice-wise Grad-CAM maps are produced, visually highlighting the anatomical cues that guide the model’s decisions. Results: Experiments conducted on the newly curated LungCA dataset comprising 767 patients demonstrate that the model achieves 97.40% accuracy in the Turkish (TR) reporting scenario and 94.81% accuracy in the English (EN) scenario, alongside strong alignment with human-written reports in BLEU, ROUGE, METEOR, and CIDEr metrics. Conclusions: The findings demonstrate that the proposed multi-slice transformer framework achieves robust performance in both classification and radiology report generation, enhances transparency throughout the decision-making process, and provides a robust artificial intelligence solution capable of effectively supporting clinical workflows in lung cancer assessment.
KATAR et al. (Wed,) studied this question.