3D Deeply Supervised Network for Automatic Liver Segmentation from CT Volumes

Abstract

Automatic liver segmentation from CT volumes is a crucial prerequisite yet challenging task for computer-aided hepatic disease diagnosis and treatment. In this paper, we present a novel 3D deeply supervised network (3D DSN) to address this challenging task. The proposed 3D DSN takes advantage of a fully convolutional architecture which performs efficient end-to-end learning and inference. More importantly, we introduce a deep supervision mechanism during the learning process to combat potential optimization difficulties, and thus the model can acquire a much faster convergence rate and more powerful discrimination capability. On top of the high-quality score map produced by the 3D DSN, a conditional random field model is further employed to obtain refined segmentation results. We evaluated our framework on the public MICCAI-SLiver07 dataset. Extensive experiments demonstrated that our method achieves competitive segmentation results to state-of-the-art approaches with a much faster processing speed.

Segmentation of Organs at Risk in thoracic CT images using a SharpMask architecture and Conditional Random Fields

“…This architecture allows the network to learn the ω m parameters and the compatibility function μ . Typically the Potts model is used [9]; however, it has the limitation of giving a fixed penalty to similar voxels with different label assignments. Learning the compatibility function gives the flexibility of penalizing differently the organ assignments to voxel pairs according to their appearance and position, effectively accounting for relationships between them [7].…”

“…This is due to the ability of the network to combine low level features (from early layers) with high level features (from deep layers), reducing some of the issues that arise due to loss of resolution given by the use of pooling layers in semantic segmentation. Additionally, with the aim of overcoming the implicit coarseness given by FCN architecture, and to further enforce spatial relationships between the organs, we used Conditional Random Fields (CRF) as refinement step where different from other works [9], we used the CRFas- RNN [7] architecture since it allows the operation to be part of the network making the full system trainable end to end. Extensive experiments demonstrated that our method outperforms regular FCN architectures, their combination with CRF, and atlas methods like patch-based label fusion.…”

Joint Segmentation of Multiple Thoracic Organs in CT Images with Two Collaborative Deep Architectures

“…Both the first and second architectures use large filters (i.e., 7 × 7, or 7 × 7 × 7), as large filters have been shown beneficial for CT segmentation [2]. Both 2D and 3D settings are implemented in our study.…”

Fully automated esophagus segmentation with a hierarchical deep learning approach

“…We use rather large kernel size (7×7) for the filters, in accordance with other works in CT images [18]. Stochastic gradient descent (SGD) is used with a learning rate of 0.1 which was decreased by a factor of ten every 20 epochs and initialized by Xavier initialization [19] for all the weights in the networks.…”