Exploring pretrained encoders for lung nodule segmentation task using LIDC-IDRI dataset

Deep learning has become ubiquitous in the field of computer vision for tasks such as image classification and segmentation. A Computer-Aided Diagnostic (CAD) system for lung cancer detection and diagnosis works by identifying lung nodules and characterizing the same. Transfer learning allows for pre-trained weights to be ported from one model to another. Replacement of pre-trained encoders in encoder-decoder networks opens up the number of possibilities of such networks and motivates us to check the possibility of each combination for a segmentation task of interest. This paper reports the experiments carried out using such combinations and presents the various observations as a result of the experiments for the nodule segmentation task on the LIDC-IDRI dataset. This work also examines the effect of network parameters on some of the deep learning semantic segmentation architectures in the context of the lung cancer dataset, LIDC-IDRI. The efficient network architecture, based on observations, is determined to be UNet with the backbone architecture, Efficientnet-b3 trained on the ImageNet dataset. This specific network presents an IoU score of 0.59 on the training dataset and 0.45 on the validation dataset. The architectures were compared and analyzed in terms of the time and space taken as well.