This project tackles the challenge of accurately segmenting lungs from 3D CT scans using deep learning, employing a 5-level 3D U-Net architecture built with TensorFlow/Keras and trained on the LCTSC dataset from The Cancer Imaging Archive. To enhance model generalization and mitigate overfitting, extensive data augmentation techniques—such as random flips, rotations, and elastic deformations—were implemented alongside dropout and L2 weight decay regularization. The resulting system achieves reliable lung segmentation across diverse patient anatomies, demonstrating the effectiveness of combining a deep network with robust training strategies.

You can find the complete code and further details on GitHub.