Lung nodule segmentation in CT scans using 3d u-net models with inception and resnet architectures

Abstract

Lung cancer is a significant contributor to global cancer-related fatalities. Utilizing deep learning techniques has the potential to enhance the detection rate of lung nodules in chest computed tomography (CT) scans, facilitating effective diagnosis and treatment planning. Moreover, given the three-dimensional (3D) nature of thoracic CT scans, 3D CNNs can effectively capture volumetric information embedded in the images. This study explores lung nodule segmentation using various 3D U-Net architectures: Standard 3D U-Net (Model-I), ResNet 3D U-Net (Model-II), Inception 3D U-Net (Model-III), and Inception-ResNet 3D U-Net (Model-IV); optimized with a combination of Dice and Focal Loss. These models were evaluated on an L1-patient test dataset. Model-II outperforms all the models, achieving an average dice similarity coefficient (DSC_avg)) of 0.84±0.06 and an average Jaccard index (IoU_avg) of 0.73±0.09 on the test dataset. Moreover, this study identifies compromised predictions generated by the models and their associated characteristics. Addressing the issues with superior approaches is a crucial aspect to be considered in future work.