Room: Exhibit Hall
Purpose: The aim of this work is to develop an end-to-end deep learning framework for liver segmentation, which helps determine the shape and size of the lesions and localize them.
Methods: The proposed semantic segmentation network takes three-dimensional CT volume of 448x448x5 voxels as input, and produces an automatically labeled image corresponding to the central slice. Our network consists of five convolutional layers and four up-convolutional layers as the encoder and decoder parts, respectively. To maximize the accuracy of our results, we adopt the methods of the densely connected segmentation network: the convolutional layers concatenate input and output feature maps of each convolution block for the following network layers. For training the network, we adopt the Liver Tumor Segmentation (LiTS) dataset. After the training, our model is evaluated with average pixel-level sensitivity and Intersection of Union (IOU) as validation metric.
Results: When we compare the proposed method with the existing models highly ranked in the benchmark, the proposed network outperforms the existing algorithms in the benchmark. Especially, our method shows a higher sensitivity in finding liver tumor area compared to the existing algorithms.
Conclusion: In this work, we propose the semantic segmentation network with the concatenation of input and output feature maps of each convolution block for the following network layers. The results show that our method has a higher sensitivity in finding liver tumor compared to the existing algorithms.