Automatic segmentation of healthy liver in abdominal computed tomography scans

Abstract

Segmentation is the process of delineating regions of interest and this process is applied to medical scans to help with diagnosis of diseases as well as treatment planning and monitoring. At the date of writing this work, segmentation is primarily carried out manually by medical professionals, which adds a substantial workload. Convolutional Encoder-Decoders (CEDs) currently dominate the medical image automatic segmentation field and many have produced satisfactory results, given the limited availability of training data. This work explores literature of some of these implementations and goes into detail about a state-of-the-art model called v16pUNet1.1C, which is an architecture based on VGG16, UNet and the Cascade Framework. The Combined Healthy Abdominal Organ Segmentation (CHAOS) Challenge database and its Task 2 framework are used to replicate and verify the state-of-the-art implementation. A modification of the architecture of v16pUNet1.1C was carried out with the purpose of increasing the performance. Modifications were also performed on the learning rate, context connections and the cascade framework, however, none seemed to lead to an increase in mean score performance, although they did narrow the interquartile range, which is a success in its own merit. The modified model, called v16pUNet1.1D, managed to achieve a mean score of 85.92, just 0.06 points shy from first place in Task 2 of the CHAOS Challenge.