U-Net and Mean-Shift Histogram for Efficient Liver Segmentation from CT Images
Boonnatee Sakboonyara, Pinyo Taeprasartsit · 2019
Nowadays, 3D computed tomography (CT) data is commonly employed to diagnose liver cancer or to examine liver condition. This work proposes a method that can quickly find a location of the biggest CT slice of the liver. We also extend this method to demonstrate that it can work as an efficient algorithm for liver segmentation. The proposed method is based on a U-Net, fully convolutional neural network, to roughly segment the liver and employs a mean-shift clustering algorithm to enhance liver localization accuracy. This addition of mean-shift clustering on histogram data after rough segmentation by U-Net prevents oversegmentation of the liver and significantly improves the accuracy of a method solely relying on a U-Net. A novel image-enhancement technique based on statistical thresholding is also introduced to further increase accuracy. This hybrid method processed each slice in 0.35 seconds on average (75 seconds per image) and its median accuracy measured in Dice similarity index in 5-fold cross-validation was 95%.