Article
LUNG CANCER SEGMENTATION USING ML & DL
Lung cancer kills more people each year than almost any other cancer, and the main reason is late diagnosis. Once a tumour reaches an advanced stage, treatment options shrink dramatically. Screening programs that use Computed Tomography (CT) imaging can catch nodules early, but reading hundreds of CT slices per patient is slow, and two radiologists looking at the same scan often disagree. This paper describes an automated segmentation system built on a deep U-Net architecture that takes raw CT slices as input and returns pixel-level nodule masks without any human in the loop. The encoder half of the network compresses the image into a compact feature representation while capturing broad context; the decoder half reconstructs a full-resolution map; and skip connections carry fine spatial detail directly from each encoder stage to its matching decoder stage so that thin nodule boundaries are not smeared during upsampling. The model was trained on the publicly available LIDC-IDRI dataset using an Adam optimizer with binary cross-entropy loss. Segmentation quality was measured with the Dice Similarity Coefficient and Intersection over Union, both of which reward precise overlap rather than raw pixel accuracy. On the held-out test split the system reached a Dice score of approximately 0.89 and an IoU of approximately 0.82, outperforming classical regiongrowing methods and shallow convolutional baselines. The results suggest this pipeline can meaningfully support radiologists in routine screening workflows.
Full Text Attachment
