DEEP LEARNING MODELS FOR EARLY DETECTION OF LUNG ADENOCARCINOMA

Abstract

The growth of cancer cells often begins as small nodules or distorted growth of bronchioles or alveoli lining. By causing abnormal growth of cells that normalize the oxygen exchange mechanism, these cells thin the lining and constrict the air passages. During the initial phase, the patients might present with mild, persistent cough, tightness in the chest or dyspnea following exertion. The reason why the issue of early diagnosis of a lung adenocarcinoma which is a major cause of cancer mortality in the world continues to be a challenge in clinical practices is because of the insidious nature and gradualist of signs and symptoms at early stages. In this study, the researcher will propose an effective diagnostic model based on deep learning that incorporates CNN, DenseNet-121, and Swin Transformer to achieve a higher detection rate and explain ability. The preprocessing of the LUNA16 CT data was performed using Hounsfield Unit clipping (1000 to 400), resampling, segmentation and 3D patch extraction. All of the models were trained on Focal and Dice losses based on the Adam optimizer and validated through five-fold cross-validation. Spatial details and nodule texture were captured in the CNN, and the flow of features and information in DenseNet-121 was supported. Utilizing hierarchical attention, the Swin Transformer, which models long-term dependencies and contextual associations of tissues, was able to outperform other architectures. The CNN, DenseNet-121, and Swin Transformer achieved 91.2%, 94.8 and 98.1 per cent accuracy, respectively, which substantiates significant gains in the accuracy of classification and sensitivity. Malignant and benign cases were discriminated consistently with the help of confusion matrices and ROC curves.