Iterative fully convolutional neural networks for automatic vertebra segmentation and identification (1804.04383v3)

Abstract: Precise segmentation and anatomical identification of the vertebrae provides the basis for automatic analysis of the spine, such as detection of vertebral compression fractures or other abnormalities. Most dedicated spine CT and MR scans as well as scans of the chest, abdomen or neck cover only part of the spine. Segmentation and identification should therefore not rely on the visibility of certain vertebrae or a certain number of vertebrae. We propose an iterative instance segmentation approach that uses a fully convolutional neural network to segment and label vertebrae one after the other, independently of the number of visible vertebrae. This instance-by-instance segmentation is enabled by combining the network with a memory component that retains information about already segmented vertebrae. The network iteratively analyzes image patches, using information from both image and memory to search for the next vertebra. To efficiently traverse the image, we include the prior knowledge that the vertebrae are always located next to each other, which is used to follow the vertebral column. This method was evaluated with five diverse datasets, including multiple modalities (CT and MR), various fields of view and coverages of different sections of the spine, and a particularly challenging set of low-dose chest CT scans. The proposed iterative segmentation method compares favorably with state-of-the-art methods and is fast, flexible and generalizable.

• The paper presents an iterative FCN approach that segments one vertebra at a time using a memory component for improved sequential analysis.
• It achieves high precision with an average Dice score of 94.9% and a symmetric surface distance of 0.2 mm, ensuring accurate anatomical labeling.
• The method adapts robustly to various imaging modalities, enhancing automatic spine analysis and clinical decision-making.

Iterative Fully Convolutional Neural Networks for Automatic Vertebra Segmentation and Identification

The paper focuses on the automatic segmentation and identification of vertebrae using iterative fully convolutional neural networks (FCNs). This work is seminal in removing the dependency on the number of visible vertebrae and specific vertebrae visibility, an issue prevalent due to varying imaging fields across CT and MR scans. The approach is critical for enabling automatic spine analysis procedures such as detecting vertebral compression fractures.

Methodology Overview

The authors propose an iterative instance segmentation approach, deploying an FCN to process each vertebra independently. The network segments one vertebra at a time, employing a memory component that remembers information about previously segmented vertebrae. This allows the network to focus on locating and identifying new vertebrae one-by-one. A unique feature of this method is its reliance on the collinearity of vertebrae for traversal of the image.

The network performs multiple tasks:

• Segmentation: Achieved through a binary classification at the voxel level in a patch.
• Anatomical Labeling: Estimates vertebra labels using a maximum likelihood approach which considers the likelihood of anatomical sequences.
• Completeness Classification: Distinguishes between fully and partially visible vertebrae to prevent incorrect data inclusion in further analyses.

Evaluation and Results

The method was assessed using five distinct datasets with varying modalities, slice thicknesses, and anatomical coverage, strengthening claims about its robustness. Its efficacy is evident in achieving an average Dice score of 94.9% and average absolute symmetric surface distance of 0.2 mm. Vertebral recognition accuracy was 93%, with a 97% accuracy in determining vertebral completeness.

Implications and Future Directions

The proposed approach compares favorably with current state-of-the-art methods due to its flexibility and adaptability across various imaging modalities and conditions. Practically, this robustness is transformative for clinical settings where images are not typically tailored for spinal analysis. Theoretically, this work contributes to the understanding of iterative segmentation strategies and the capability of neural networks to perform multi-task learning efficiently.

Moving forward, the integration of more sophisticated models addressing irregularities in vertebral counts and significant deformations directly within the network framework could enhance the approach's capabilities. Advances in hardware will allow exploration of deeper network architectures, potentially improving segmentation precision and speed.

Overall, the paper represents a significant contribution to vertebra segmentation using machine learning techniques, expanding the potential of FCNs in the medical imaging domain.