Automatic Liver Segmentation Using an Adversarial Image-to-Image Network (1707.08037v1)

Abstract: 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.

• The paper introduces a deep image-to-image network combined with adversarial training to accurately segment livers in 3D CT scans.
• It employs a convolutional encoder-decoder with multi-level feature concatenation and deep supervision, achieving a Dice coefficient of 0.95 and a mean ASD of 1.90 mm.
• The method's fast and robust performance underlines its potential clinical application in diagnosis, treatment planning, and postoperative assessment.

Automatic Liver Segmentation Using an Adversarial Image-to-Image Network

The paper presents an advanced method for automatic liver segmentation from 3D computed tomography (CT) volumes by leveraging an adversarial image-to-image network. The authors introduce a deep image-to-image network (DI2IN) employed as a generator in conjunction with an adversarial network to perform this segmentation task. This approach targets a critical medical imaging challenge due to complexities associated with liver segmentation, including fuzzy boundaries and variations in liver appearance across different patients and imaging protocols.

Method Overview

The core methodology hinges on a convolutional neural network (CNN) architecture, specifically a convolutional encoder-decoder structure, to perform voxel-wise binary classification, predicting the probability of each voxel belonging to the liver region. The network integrates multi-level feature concatenation and deep supervision to improve segmentation accuracy. The convolutional kernels are designed with a size of $3\times3\times3$, and the model employs bilinear interpolation in upscaling layers to enhance learning efficiency and reduce computational load.

To refine DI2IN's predictions, an adversarial framework is employed. This consists of a generator (DI2IN) and a discriminator network trained via adversarial training, where the generator's objective is to produce liver segmentations indistinguishable from true annotations. The discriminator assists the generator by identifying differences between the segmentation outputs and ground truth, ultimately refining the segmentation accuracy.

Experimental Evaluation

The proposed method is validated on a substantial dataset of over 1000 annotated CT volumes, surpassing previous benchmarks in terms of dataset size for training liver segmentation networks. This abundance of training data facilitates the network's ability to generalize across numerous variables, such as contrast levels, resolutions, and positions, as well as demographic and pathological diversity.

Their algorithm's performance showcases superior precision over both classical learning-based methods and prior CNN-based techniques. The quantitative results, highlighted by metrics such as average symmetric surface distance (ASD) and Dice coefficients, reflect significant improvements. The DI2IN-AN achieved a mean ASD of 1.90 mm and an average Dice coefficient of 0.95, representing the most accurate performance among compared methodologies.

Implications and Future Directions

This innovative approach has solid implications for clinical practices involving medical imaging. Accurate liver segmentation enables enhanced diagnosis, treatment planning, and postoperative assessments. The adversarial network's contribution to precision further reinforces its potential applicability in broader medical imaging segmentation tasks. The remarkable computational efficiency, taking under one second to segment a CT volume, underscores its feasibility for routine clinical application.

Looking forward, this work lays a foundation for further research into adversarial training frameworks in medical image analysis. Potential advancements could involve exploring alternative architectures and loss functions to further refine segmentation outputs. Additionally, the approach could be tested on new imaging modalities, such as MRI, to evaluate its versatility across different medical imaging technologies.

In sum, this paper introduces a robust automated liver segmentation method that harnesses adversarial learning to attain high segmentation accuracy and computational efficiency, setting a new standard in the field of medical image segmentation using deep learning models.