Source-linked AI summary
3D Deeply Supervised Network for Automatic Liver Segmentation from CT Volumes
Qi Dou, Hao Chen, Yueming Jin, Lequan Yu, Jing Qin, Pheng-Ann Heng
TL;DR
Automatic liver segmentation from CT volumes is important for hepatic diagnosis and treatment planning but is challenged by anatomical variation, low contrast, and pathology. The paper proposes a fully convolutional 3D deeply supervised network with CRF contour refinement, achieving competitive segmentation against state-of-the-art approaches and faster processing on MICCAI-SLiver07.
Problem
Automatic CT liver segmentation is important but difficult because of shape variation, low contrast with adjacent organs, and diverse pathologies, while volumetric context and high-level feature extraction remain open problems.
Source: §1 · pp. 1–2Method
The method combines a fully convolutional 3D DSN with hidden-layer deep supervision and fully connected CRF contour refinement.
Source: §1 · p. 2, §2.2 · p. 4, §2.3 · p. 5Results
VOE of 5.42% and AvgD of 0.79mm were achieved on the testing set, with processing taking about 1.5 minutes per subject.
Source: §3 · p. 7Takeaways & Limitations
Abstract
from arXiv · showhide
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.
1 Introduction
Automatic liver segmentation from CT volumes is important for hepatic diagnosis and treatment planning but remains difficult because of anatomical variation, weak boundaries, and pathology. The paper proposes a 3D deeply supervised network to address unresolved challenges in volumetric context modeling and high-level feature extraction.
- Accurate liver segmentation supports computer-aided hepatic disease diagnosis and treatment planning.Source: §1 · p. 1
- Automatic segmentation is motivated by manual annotation that is tedious, error-prone, and time-consuming.Source: §1 · p. 1
- CT liver segmentation is challenging because of inter-patient shape variation, low contrast with adjacent organs, and varied pathologies.Source: §1 · p. 1
- Leveraging volumetric contextual information and extracting powerful high-level representations for automatic liver segmentation remains an open problem.Source: §1 · p. 2
- The proposed 3D DSN uses fully convolutional end-to-end learning, deep supervision, and CRF contour refinement for automatic liver segmentation.Source: §1 · p. 2
2 Method
The method combines an end-to-end 3D fully convolutional architecture with hidden-layer supervision and CRF-based contour refinement. Deep supervision addresses difficult optimization, while the CRF refines ambiguous boundaries using network predictions and image information.
2.1 End-to-end 3D Fully Convolutional Architecture
- The 3D DSN uses an 11-layer 3D architecture with convolution, pooling, deconvolution, softmax, and supervision branches at the third and sixth layers.Source: §2 · p. 2, §2.1 · p. 2
- Deconvolutional layers reconstruct dense predictions by up-scaling coarse feature volumes with trainable in-network kernels.Source: §2.1 · p. 3
- The network inputs a large volume and directly outputs an equal-sized prediction score map through fully convolutional end-to-end learning and inference.Source: §2.1 · p. 3
2.2 Deep Supervision for Learning Process
- Per-voxel binary classification learning faces vanishing gradients, which can slow convergence and reduce discrimination capability, especially in 3D.Source: §2.2 · p. 3
- Deep supervision adds auxiliary dense-prediction losses from hidden layers, with simultaneous back-propagation from branch and final outputs.Source: §2.2 · p. 4
2.3 Contour Refinement with CRF
- A fully connected CRF refines ambiguous contours using unary potentials initialized from weighted network predictions and pairwise terms incorporating appearance and smoothness.Source: §2.3 · pp. 4–5
3 Experiments
Experiments evaluated the 3D DSN on MICCAI-SLiver07 using training-set leave-one-out analysis and held-out testing comparisons. The method outperformed 3D CNN and achieved strong accuracy with substantially faster processing than competing shape-modeling approaches.
Dataset and setup
- MICCAI-SLiver07 contains 30 contrast-enhanced CT scans, with 20 for training and 10 for testing.The implementation was trained from scratch with scheduled learning-rate and deep-supervision weight decay.Source: §3 · p. 5
Training-set evaluation
- Training-set leave-one-out evaluation used VOE, VD, AvgD, RMSD, and MaxD, where lower absolute values indicate better segmentation.Table 1 evaluated the proposed methods under different settings using these five measures.Source: §3 · p. 6, Table 1 · p. 6
- 3D DSN yielded superior training-set results to 3D CNN, indicating benefits from deep supervision for optimization and discrimination.The comparison was reported across the evaluation measures in Table 1.Source: §3 · p. 6, Table 1 · p. 6
Qualitative results
- The method delineated livers despite low contrast, inter-patient shape variation, and internal pathologies with abnormal appearance.Figure 3 shows ground truth in green and predictions in red across transverse, sagittal, and coronal views.Source: §3 · p. 6, Figure 3 · p. 7
Testing-set evaluation
- 5.42% VOE and 0.79mm AvgD were achieved on the held-out testing set against top-ranking teams and published state-of-the-art approaches.These were identified as the two most important and commonly used liver-segmentation metrics.Source: §3 · p. 7, Table 2 · p. 7
- 1.5 mins per subject was the framework’s processing time, including 5s for 3D DSN and 87s for CRF.The paper reports this as faster than competing shape-modeling approaches, which commonly took several minutes.Source: §3 · p. 7
4 Conclusion
The paper presents a 3D deeply supervised network for automatic liver segmentation in abnormal CT volumes, followed by CRF-based contour refinement. Results on SLiver07 were promising and substantially faster, while the approach is described as extensible to other volumetric segmentation applications with limited training data.
4 Conclusion
- The 3D DSN generates high-quality score maps, and a conditional random field refines their contours.This combines the network’s volumetric segmentation output with a subsequent graphical-model refinement stage.Source: §4 · p. 8
- On SLiver07, the method achieved promising segmentation results with much faster processing speed.The conclusion summarizes the reported dataset-level outcome without specifying an additional metric.Source: §4 · p. 8
- The deep-learning method is described as general and readily extendable to other medical volumetric segmentation applications with limited training data.This is the paper’s stated broader applicability claim.Source: §4 · p. 8