Vertebrae Identification and Localization Utilizing Fully Convolutional Networks and a Hidden Markov Model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Automated vertebrae identification and localization in spinal CT imaging is a complex task requiring both image feature extraction and sequence modeling.
• Existing methods often struggle with the long sequence detection and spatial interdependence inherent in 3D medical images.

Purpose of the Study:

• To develop a robust and accurate framework for automated vertebrae identification and localization in 3D spinal CT images.
• To leverage fully convolutional neural networks (FCNs) for simultaneous feature extraction and sequence modeling in this hybrid task.

Main Methods:

• A 3D fully convolutional neural network (FCN) was trained end-to-end at the spine level to capture long-range contextual information.
• Input downscaling and an auxiliary FCN were used to manage computational load and preserve image details.
• A hidden Markov model (HMM) was integrated to impose explicit spatial and sequential constraints for enhanced robustness.

Main Results:

• The proposed framework achieved an identification rate (within 20 mm) of 94.67% on the test set.
• A mean identification rate of 87.97% and a mean error distance of 2.56 mm were recorded.
• The results represent the highest performance reported on the MICCAI 2014 Vertebrae Localization and Identification dataset.

Conclusions:

• The developed 3D FCN framework effectively integrates local image details and global patterns for accurate vertebrae identification and localization.
• The incorporation of HMM significantly enhances the robustness and interpretability of the network's outputs.
• This approach sets a new benchmark for automated spinal CT image analysis, offering potential for improved clinical diagnostics.