Journal of Pioneering Artificial Intelligence Research

Abstract

Early stage identification is crucial for improving patient outcomes because lung cancer is one of the leading causes of high death rates globally. A deep learning-based method for automatically segmenting lung tumors from computed tomography (CT) images is presented in this work. Following routine preprocessing and data augmentation, a U-Net architecture was trained using the LIDC-IDRI dataset and the Kaggle rasoulisaeid/lung-cancer-segment dataset. A Streamlit application is used to install the final model (lung-cancer-segmenter.pth), which enables users to upload CT scans and see segmented tumor locations in real time. A Dice coefficient of 0.0966, IoU of 0.0507, accuracy of 0.9925, and sensitivity of 0.7269 were obtained from evaluation on 3,154 test samples. Due to class imbalance and the tiny size of tumor areas, the model exhibits great sensitivity but low precision. The method serves as a useful computer-aided tumor localization tool, and future development will concentrate on increasing segmentation accuracy and incorporating stage predication. The novelty of this work lies in the development of an end -to – end and lightweight lung tumor segmentation framework that integrates preprocessing, U-net based segmentation, and real-time visualization within a single Streamlit based application. Unlike conventional studies that only focus solely on segmentation accuracy and precision, the proposed approach emphasizes practically usability by enabling the interactive and responsive lung tumor visualization and classification that provides a foundation for the stage prediction, thereby supporting early clinical assessment.