Advancing Lung Cancer Detection Using Optimized Fine Tuning Data-Efficient Image Transformers

According to estimates from the World Health Organization (WHO) in 2019, cancer is the first or second leading cause of death before the age of 70 years in 112 of 183 countries (Sung et al. in CA Cancer J Clin 71(3):209–249, 1). The localized cases have a five-year survival rate of 56% though. In metastatic disease, there may be only 5% of evidence of the significance of early diagnosis. Although low-dose CT screening has achieved a 20% reduction in mortality (Sung et al. 2021), the heaviness of the workload and lack of consistency in reading the scans are still serious challenges to broad adoption. Existing CNN-based CAD systems are useful, however, may often miss subtle but clinically significant features like ground-glass opacities and spiculated margins due to their localized receptive fields. In this paper, we unleash a tweaked pictorial converter (DeiT) for sorting lung knots from slice-based imagery, shattering convolutional chains to embrace worldwide pattern weaving. Rigorous trials on dual open archives—the Iraqi cancer hub’s IQ-OTH/NCCD and the consortium’s LIDC-IDRI—yielded stellar scores: 99.54% precision on the former and 95.64% on the latter. The findings with the DeiT model are either on par or better than CNN-based models, indicating the possibility of transformer architectures to one day be the new standard of medical image classification.