Source-linked AI summary
Automatic Liver Segmentation Using an Adversarial Image-to-Image Network
Dong Yang, Daguang Xu, S. Kevin Zhou, Bogdan Georgescu, Mingqing Chen, Sasa Grbic, Dimitris Metaxas, Dorin Comaniciu
TL;DR
Automatic liver segmentation is clinically important but difficult because liver appearance and boundaries vary in 3D medical images. The paper combines DI2IN with adversarial training and more than 1000 diverse annotated CT volumes, reporting improved segmentation quality and faster processing than prevailing methods.
Problem
Automatic liver segmentation is difficult because medical images contain complex backgrounds, fuzzy boundaries, and varied liver appearance, despite its clinical importance.
Method
DI2IN performs voxel-wise liver segmentation with an encoder-decoder, feature concatenation, and deep supervision, while adversarial training distinguishes predictions from ground truth.
Results
The proposed method achieves the best ASD and dice performance among five compared methods, while DI2IN shows a 20% improvement in ASD over examined prevailing networks.
Takeaways & Limitations
Training on more than 1000 annotated 3D CT volumes improves CNN performance by a large margin, while the proposed algorithm provides good segmentation quality and faster processing speed.
Abstract
from arXiv · showhide
Automatic liver segmentation in 3D medical images is essential in many clinical applications, such as pathological diagnosis of hepatic diseases, surgical planning, and postoperative assessment. However, it is still a very challenging task due to the complex background, fuzzy boundary, and various appearance of liver. In this paper, we propose an automatic and efficient algorithm to segment liver from 3D CT volumes. A deep image-to-image network (DI2IN) is first deployed to generate the liver segmentation, employing a convolutional encoder-decoder architecture combined with multi-level feature concatenation and deep supervision. Then an adversarial network is utilized during training process to discriminate the output of DI2IN from ground truth, which further boosts the performance of DI2IN. The proposed method is trained on an annotated dataset of 1000 CT volumes with various different scanning protocols (e.g., contrast and non-contrast, various resolution and position) and large variations in populations (e.g., ages and pathology). Our approach outperforms the state-of-the-art solutions in terms of segmentation accuracy and computing efficiency.
1 Introduction
Automatic liver segmentation is clinically important but difficult because 3D medical images contain complex backgrounds, fuzzy boundaries, and varied liver appearance. The paper proposes DI2IN-AN, combining a deep image-to-image segmenter with adversarial training and a large, diverse CT dataset.
- 3D liver segmentation supports pathological diagnosis, surgical planning, and postoperative assessment.
- Automatic segmentation remains challenging because of complex backgrounds, fuzzy boundaries, and varied liver appearance.
- Existing approaches include non-learning-based methods and learning-based classifiers using hand-crafted features.
- CNN approaches use end-to-end segmentation with refinement methods such as conditional random fields or graph cuts, but these refinements can be time-consuming and leak across low-contrast boundaries.
- DI2IN-AN uses DI2IN as a generator, combining convolutional encoder-decoder layers, multi-level feature concatenation, deep supervision, and an adversarial discriminator.
- The method leverages more than 1000 annotated CT volumes spanning scanning protocols, resolutions, positions, ages, and pathology.
2 Methodology
DI2IN performs voxel-wise liver segmentation with a 3D convolutional encoder-decoder, multi-level feature concatenation, and branch-level supervision. Adversarial training adds a discriminator-driven objective that guides DI2IN predictions toward ground-truth labels during training, while inference uses the generator alone.
- Deep Image-to-Image Network (DI2IN): DI2IN treats liver segmentation as voxel-wise binary classification and maps entire 3D CT volumes to liver-probability maps.Its symmetric architecture uses a convolutional encoder-decoder built from 3D convolutional and bilinear upscaling layers.
- Deep Image-to-Image Network (DI2IN): Feature-layer concatenation bridges encoder and decoder blocks, combining advanced and low-level features to provide local and global contextual information.The concatenated features are passed to the next convolutional layer.
- Adversarial Training: The adversarial scheme trains a generator to produce segmentation predictions while a discriminator distinguishes predictions from ground truth.The discriminator is assigned ground-truth and prediction labels during training and is optimized through an adversarial objective.
- Deep Image-to-Image Network (DI2IN): Branch outputs receive voxel-wise binary cross-entropy supervision, and the total loss weights the final-output and branch loss terms.Each branch is upscaled to the input-image size before its loss is calculated against the ground truth.
- Adversarial Training: The generator loss combines conventional voxel-wise segmentation loss with an adversarial component that encourages discriminator confusion between predictions and ground truth.During discriminator training, its loss gradient is propagated to update DI2IN parameters.
- Adversarial Training: Generator and discriminator updates alternate for multiple training steps, after which the adversarial network is removed and the generator performs inference alone.The training procedure repeatedly generates predictions and updates the discriminator and generator in separate loops.
3 Experiments
Experiments use a small public-dataset comparison and an independent 50-volume test to assess segmentation accuracy and efficiency. Results favor DI2IN-AN, while larger training data improves CNN performance.
- Dataset: The collected dataset includes more than 1000 CT volumes, addressing the limited scale of public liver-segmentation datasets.MICCAI-SLiver07, for example, contains 20 training and 10 testing volumes, all contrast enhanced.
- Results: DI2IN-AN achieves the best performance in both evaluation metrics among the compared methods.The comparison also reports that deep-learning methods outperform the classic hand-crafted-feature algorithm.
- Results: 0.23mm lower mean ASD and 3.84mm lower maximum ASD result when DI2IN is trained on 1000+ rather than 400 labelled cases.The passage attributes these changes to using more training data.
- Evaluation: Table 1 compares five methods on 50 unseen CT data using average symmetric surface distance and Dice coefficients.DI2IN and DI2IN-AN were trained with either 400 or 1000+ cases, enabling comparisons of training-set size and adversarial training.
- Efficiency: DI2IN achieves a 20% improvement in average symmetric surface distance over the examined prevailing networks.The paper also characterizes DI2IN as faster at inference than often time-consuming CNN-based methods.
4 Conclusion
The paper presents an adversarial image-to-image liver-segmentation algorithm trained on more than 1000 annotated 3D CT volumes. It reports good segmentation quality and faster processing speed, with large-dataset training improving CNN performance.
- The proposed algorithm uses an adversarial image-to-image network for automatic liver segmentation.
- The method achieves good segmentation quality and faster processing speed.
- Training on more than 1000 annotated 3D CT volumes can improve CNN performance by a large margin.