Abstract In the intricate domain of lung cancer diagnostics, this research presents a groundbreaking for the early detection of lung cancer at the cellular level which remains a critical challenge due to the subtle morphological differences among its subtypes. This study introduces a hybridized deep‐learning framework that combines the global feature expertise of ResNet-50 with the spatial-attention capabilities of Attention U-Net to analyse microscopic images of individual lung cells. A comprehensive image‐processing pipeline featuring Contrast Limited Adaptive Histogram Equalization (CLAHE), median‐filter for denoising, Otsu’s adaptive thresholding for feature extraction and targeted grey- midtone lightening to amplify diagnostically irrelevant textures and suppresses artifacts, which boosts the signal-to-noise ratio by 23%. Using a balanced dataset of 4,650 grayscale images (1,500 per subtype) and enriched extensive image augmentations the model learned robust representations across adenocarcinoma, neuroendocrine carcinoma, and squamous cell carcinoma. On a 25% hold-out test set, it achieved 99.85% overall accuracy, with precision, recall, and F1-scores all exceeding 0.99 (for neuroendocrine carcinoma). Five-fold stratified cross-validation confirmed this performance (mean accuracy 99.69% ± 0.16%), demonstrating exceptional consistency and minimal variance. By detecting cancer at its very inception in single‐cell images, this approach paves the way for ultra early diagnostics and personalized treatment planning in clinical practice at the very initially cellular level.
H. Parveen Sultana (Mon,) studied this question.