TL;DR
This paper introduces a deep convolutional neural network (DCNN) that can accurately measure vertebral body dimensions and detect insufficiency fractures from lumbar spine MRIs. The DCNN demonstrates performance comparable to experienced radiologists, offering a promising tool for automated analysis in clinical practice.
What Problem Does This Paper Address?
Vertebral body fractures, particularly insufficiency fractures common in older adults with osteoporosis, pose a diagnostic challenge. Accurate measurement of vertebral body dimensions (height, angle, concavity) is crucial for assessing fracture severity and guiding treatment. While manual measurement is time-consuming and subject to inter-observer variability, and existing AI approaches have focused on specific tasks or imaging modalities, there's a need for a comprehensive, automated solution that can perform both precise measurements and fracture detection on heterogeneous lumbar spine MRI data. This paper aims to fill that gap by developing and validating a DCNN for these dual tasks.
Methods
The core of this work is a Deep Convolutional Neural Network (DCNN) designed for end-to-end analysis of lumbar spine MRIs. The DCNN's pipeline involves three main stages:
- Pre-processing: This step standardizes the input MRI images. It selects sagittal T2-weighted sequences, rescales them to a consistent resolution (0.6 mm x 0.6 mm per voxel), and resizes them to a specific matrix size (448 x 320). To enhance distinction between vertebral bodies and surrounding tissues, intensity values are normalized to a range of 0.0 to 1.0. Only central slices are retained for further processing.
- Vertebral Body Segmentation: A U-Net-like architecture is employed for segmenting individual lumbar vertebral bodies (L1-L5). This involves an encoding path to capture image context and a symmetric decoding path to generate a binary mask of the vertebrae.
- Corner Detection and Measurements: Once segmented, the DCNN identifies four key anatomical landmarks (corners) on each vertebral body. These landmarks are then used to automatically calculate vertebral body measurements, including anterior vertebral body height (AH), posterior vertebral body height (PH), maximum extent of superior and inferior endplate concavity (SC and IC), and the vertebral angle (VA).
- Fracture Detection: Based on the derived vertebral body measurements, the DCNN classifies each vertebra as either intact or fractured, using criteria aligned with the Genant classification.
The DCNN was trained using a Dice coefficient loss function and the Adam optimizer. The implementation leveraged the TensorFlow framework.
Datasets
The study utilized a retrospective dataset of 1000 lumbar spine MRIs from 200 patients (160 with insufficiency fractures, 40 without). These images were acquired from multiple institutions (both "in-house" and "outside") using various MRI vendors (Siemens, GE, Philips) and field strengths (1.5 T and 3.0 T), creating a heterogeneous dataset.
For training, validation, and testing the DCNN, a separate dataset of 1056 lumbar spine MRIs from 2009 to 2019 was used. Of these, 79.0% were for training, 9.8% for validation, and 9.8% for testing. A subset of 37 studies was used to optimize the measurement algorithm.
Key Insights and Ideas
- Dual-Task Capability: The DCNN's ability to perform both precise vertebral body measurements and fracture detection in a single, automated pipeline is a significant advancement. This integrated approach streamlines analysis and potentially improves efficiency.
- Robustness to Heterogeneity: The DCNN's performance remained consistent across different institutions, MRI vendors, and field strengths. This is crucial for real-world clinical adoption, as it suggests the model is generalizable and not overly sensitive to variations in imaging protocols.
- End-to-End Automation: Unlike some previous AI approaches that required manual outlining or separate steps, this DCNN operates in an end-to-end fashion, from image input to measurement and classification output. This automation is key to reducing workload and potential human error.
- Agreement with Experts: The DCNN's excellent agreement with experienced radiologists for vertebral body measurements, and high diagnostic accuracy for fracture detection, validates its clinical potential. This suggests it can serve as a reliable decision support tool.
Key Results
- Vertebral Body Measurements: The DCNN achieved excellent inter-reader agreement with radiologists for anterior/posterior vertebral height and vertebral angle (ICC > 0.94). Agreement for endplate concavity was good to excellent (ICC 0.79–0.85). Bland-Altman analysis showed good agreement for height measurements, with mean differences within a few millimeters and narrow limits of agreement.
- Fracture Detection: The DCNN demonstrated high diagnostic performance for insufficiency fracture detection, with an overall sensitivity of 0.941 (95% CI: 0.903–0.968), specificity of 0.969 (95% CI: 0.954–0.980), and accuracy of 0.962 (95% CI: 0.948–0.973).
- False Positives/Negatives: The DCNN produced 24 false positives and 14 false negatives out of 1000 analyzed vertebral bodies. Many false positives were attributed to intraspongious herniation or degenerative endplate changes, while false negatives often involved subtle fractures with minimal height loss.
- Independence of Acquisition Parameters: The DCNN's diagnostic performance for fracture detection was not significantly influenced by the institution, MRI scanner vendor, or magnetic field strength.
Limitations and Future Work
- Study Cohort: The study population consisted primarily of elderly individuals due to the nature of insufficiency fractures. While the DCNN was tested on a broad age range, the generalizability to younger populations with different fracture types might be limited.
- Reference Standard: The radiologist's classification served as the reference standard, which can be subject to inter-observer variability, especially in cases with subtle findings. The study mitigated this by using a consensus reading of three experienced radiologists.
- Real-World Scenarios: The authors acknowledge that performance might be slightly lower in real-world scenarios with severe motion artifacts or scoliosis, which were exclusion criteria.
- Future Work: The authors suggest that future projects could extend this DCNN to detect multiple lumbar spine disorders and integrate it fully into clinical workflows as a decision support tool. Further research could also explore its performance on images with more complex pathologies.
Summary updated