Source-linked AI summary
Segmentation of the Proximal Femur from MR Images using Deep Convolutional Neural Networks
Cem M. Deniz, Siyuan Xiang, Spencer Hallyburton, Arakua Welbeck, James S. Babb, Stephen Honig, Kyunghyun Cho, Gregory Chang
TL;DR
Manual segmentation limits the practical use of MRI-based proximal-femur bone-quality measurements because it is time-consuming. The paper compares supervised 2D and 3D CNN architectures against expert segmentations and reports high segmentation performance for the 3D CNN, which may help bring structural MRI measurements closer to clinical practice.
Problem
Manual segmentation of proximal-femur MR images is time-consuming, limiting the practical use of MRI measurements for bone-quality assessment.
Method
The study compares supervised 2D- and 3D-convolution CNN architectures for automatic proximal-femur segmentation against expert manual segmentations.
Results
DSC = 0.940±0.054, precision = 0.946±0.024, and recall = 0.939±0.081 for the highest-performing 3D CNN.
Takeaways & Limitations
CNN-based automatic segmentation has the potential to help bring structural MRI measurements into clinical practice.
Takeaways & Limitations
The dataset consisted of only 86 subjects, and automatic hyperparameter optimization was not implemented.
Abstract
from arXiv · showhide
Magnetic resonance imaging (MRI) has been proposed as a complimentary method to measure bone quality and assess fracture risk. However, manual segmentation of MR images of bone is time-consuming, limiting the use of MRI measurements in the clinical practice. The purpose of this paper is to present an automatic proximal femur segmentation method that is based on deep convolutional neural networks (CNNs). This study had institutional review board approval and written informed consent was obtained from all subjects. A dataset of volumetric structural MR images of the proximal femur from 86 subject were manually-segmented by an expert. We performed experiments by training two different CNN architectures with multiple number of initial feature maps and layers, and tested their segmentation performance against the gold standard of manual segmentations using four-fold cross-validation. Automatic segmentation of the proximal femur achieved a high dice similarity score of 0.94$\pm$0.05 with precision = 0.95$\pm$0.02, and recall = 0.94$\pm$0.08 using a CNN architecture based on 3D convolution exceeding the performance of 2D CNNs. The high segmentation accuracy provided by CNNs has the potential to help bring the use of structural MRI measurements of bone quality into clinical practice for management of osteoporosis.
Introduction
MRI can provide three-dimensional structural information about proximal femur bone quality, but whole-femur analysis is hindered by time-consuming and variable manual segmentation. This work investigates CNN architectures for automatic segmentation against expert manual references.
- MRI can image trabecular architecture and support structural markers or finite-element modeling for proximal-femur fracture-risk assessment.
- Whole-proximal-femur analysis targets mechanical properties or strength beyond region-specific measurements, but requires expert manual segmentation.
- Manual segmentation is time-consuming for volumetric MRI and may exhibit inter-rater variability, motivating automatic whole-femur segmentation.
- Earlier automated approaches included thresholding with 3D morphology, statistical shape models, and deformable models, but their use was limited by segmentation time and robustness to femur-shape variation.
- The study compares U-net-based 2D CNN and 3D U-net architectures for automatic proximal-femur segmentation against expert manual segmentation.
Results
The 3D CNN produced the strongest proximal-femur segmentation performance and exceeded 2D CNNs on curve-based metrics. It also achieved high overlap and classification metrics, while requiring less inference time than the 2D approach.
- AUC = 0.998±0.001 and AP = 0.982±0.005 for the 3D CNN with 32 initial feature maps and 4 layers, exceeding 2D CNN performance.
- DSC = 0.940±0.054, precision = 0.946±0.024, and recall = 0.939±0.081 for the highest-performing 3D CNN.
- Post-processing improved 2D CNN segmentation DSC by 7% on average and improved precision, while average recall was not significantly affected.
- The 3D CNN modeled global proximal-femur connectivity during training, producing better delineation and avoiding additional post-processing required by the 2D CNN.
- Inference took approximately 18 seconds for 2D CNNs and 5 seconds for 3D CNNs for one subject with 48 central coronal slices.
Discussion
The study presents CNN-based automatic proximal-femur segmentation and finds improved performance with 3D convolutions. The authors view the approach as potentially bringing structural MRI measurements closer to clinical practice, while noting dataset, optimization, and architecture-scope limitations.
- CNN-based automatic segmentation may avoid expert location specifications and reduce the 1.5–2 hours of manual effort required for high-resolution volumetric datasets.
- The study extends prior proximal-femur CNN work by evaluating more subjects and comparing architectures using 2D or 3D convolution.
- 3D convolution produced better masks by directly modeling the proximal femur’s 3D connectivity, whereas 2D convolutions could leave global connectivity inadequately modeled.
- The 2D CNN used mirrored-image inference over multiple patches and averaged probability maps to generate its output segmentation mask.
- The dataset included only 86 subjects, automatic hyperparameter optimization was not implemented, and comparisons with newer architectures or alternative loss functions were beyond scope.
- The conclusion reports improved automatic proximal-femur segmentation performance using FCN and 3D convolutions, with potential relevance to clinical structural MRI use.
Methods
The study evaluates 2D and 3D U-net-based CNNs for proximal femur segmentation, using weighted training, post-processing where needed, and four-fold cross-validation against expert manual segmentations.
- CNN architectures: The 2D CNN uses a U-net with contracting and expanding paths, while the 3D CNN extends this design with volumetric convolutions, up-convolutions, and pooling.Padded operations in the 3D CNN produce an output probability map matching the input image size.
- Training: Horizontal flipping augments right- and left-hip images, while Xavier initialization and Adam optimization support CNN training.The 2D CNN also uses three consecutive slices to capture some 3D connectivity information.
- Training: The training loss is re-weighted using proximal-femur and background voxel counts to address class imbalance and emphasize trabecular-bone errors.The formulation uses Np and Nb alongside voxel-level prediction probabilities.
- Inference and post-processing: The 2D CNN uses mirrored borders, overlapping patches, and connected-volume post-processing to remove small misclassified bone clusters.The 3D CNN does not require this post-processing because 3D convolution captures volumetric connectivity.
- Study procedures: The study reports institutional review board approval and written informed consent from all subjects.The work followed regulatory and ethical guidelines for protecting human subjects.
- Evaluation: Four-fold cross-validation partitions subjects into four groups, using three groups for training and one for validation in each fold.The folds contain 21, 21, 22, and 22 subjects.
- Evaluation: Manual proximal-femur segmentations serve as ground truth, and performance is evaluated with ROC and precision-recall analyses, DSC, sensitivity/recall, and precision.DSC measures overlap between manual and automatic segmentations; ROC and precision-recall analyses support threshold selection.
- Data availability: The datasets generated and analyzed in the study are available from the corresponding author on reasonable request.
Author contributions statement
The author contributions span study conception and design, experiments, analysis, implementation, data preparation and acquisition, patient recruitment, manuscript preparation, and editing.
- Author contributions: Contributors divided responsibilities across study design, experiments, analysis, literature research, implementation, data preparation, acquisition, recruitment, and manuscript work.The statement assigns specific roles to multiple contributors, including 3D CNN implementation and data segmentation.