U-Net Segmentation of Lung Cancer CT Scans for 3D Rendering

CT scans are a common tool in lung cancer diagnosis. They help in determining whether to perform a biopsy of a detected tumor to identify if it is malicious. The work reported in this paper is aimed at the 3D Rendering of CT scans to aid in visualizing biopsies. Four variants of a U-net model were built to detect and segment the tumor area of CT scans, and classify it into one of three classes, namely Adenocarcinoma, small cell carcinoma, or squamous cell carcinoma. The implemented variants tackle problems commonly associated with machine learning applications, especially in the medical field, including insufficient annotated data, class distribution imbalance, and limited processing power. Two of the variants used focal loss as a cost-sensitive loss function to deal with imbalanced data, while the other two used categorical cross entropy. The variants were trained on segmentation masks of the scans that used two classes: background and tumor. The best performing variant achieved 99.53% accuracy, 77.27% IOU scores, and 48.60% dice coefficient score. This variant also gave the best visualization with good representation of volume and location of the tumor.