Abstract:
Automated segmentation of lung tumor from CT scan images is an essential for analyzing the progression of the lung cancer as it is one of the most widespread disease in the world. Therefore, prompt detection of malignant tumor can hence increase the possibility of patients’ survival and can help in decrease the mortality rate. In this regard, proper segmentation of suspicious lesions in computerized-tomography (CT) images is the primary step towards achieving completely automated diagnostic system for lung cancer detection. Therefore, transfer learning techniques outperform on tasks of semantic segmentation as it is an optimization that allows improved performance, saving training time and not demanding a lot of data. Models are trained on large ImageNet Datasets, and these pre-trained models are performed exceptionally well in comparison with network trained from scratch. UNET extensively used for biomedical image segmentation and have significantly improved state of the art performance. Objective of this paper is to develop an integrated architecture of two segmentation networks: MobileNetV2 and UNET, an efficient segmentation technique based on light weighted neural network developed by depth-wise separable convolutions. We trained our model on lung dataset (MSD Challenge 2018) provided by The Cancer Imaging Archive (TCIA). The suggested integrated architecture achieved dice score of 0.8793, recall of 0.8602 and precision of 0.9322 which are comparable to the results of current available techniques. Additionally, other segmentation algorithms require a lot of labeled data, while, our algorithm allows more efficient training and more generalizability to other medical image segmentation.
