Investigating the Malignancy Effect on Surrounding Tissue of Pulmonary Nodules Using Deep Learning Techniques in Lung Cancer CT Scans

Lung cancer is aggressive, with low survival rates, so early detection is crucial. Computed Tomography (CT) images are the gold standard for identifying pulmonary nodules, which are often overlooked due to their small size. This study examines the effect of nodules on the surrounding tissue to indicate cancer presence. We generate a hypothetical spread map for each image slice based on pixel distance to the nodule and extract three types of image patches: nodules, surrounding tissue without nodules, and normal tissue. These patches, and their spread maps, are fed into a regression U-Net with a customized loss function to estimate spread. Models are then trained with varying decay rates for the spread maps, followed by binary classification on predicted maps. Results show 86–91% accuracy, with a recall reaching 82.7% for surrounding tissue patches, suggesting that surrounding tissue can act as an indicator of malignant nodules.