Pulmonary Nodule Segmentation in Computed Tomography with an Encoder-Decoder Architecture

Precise pulmonary nodule segmentation from computed tomography (CT) images can contribute a lot for both clinical and research purposes while the variability of pulmonary nodules can make it difficult to perform an accurate and robust segmentation. We propose an encoder-decoder architecture which takes great advantages from convolutional neural networks and attains state-of-the-art results on the public LIDC-IDRI dataset. Our proposed method achieves an average Dice-coefficient score of 90.56% and an average surface distance of 0.13 millimeters on a five-fold cross-validation, which surpasses all the present pulmonary nodule segmentation methods on the LIDC-IDRI dataset. Besides, our method is extremely time-efficient with an average time of less than 20 milliseconds segmenting a single nodule using a modern GPU.