Purpose: Antero-posterior (AP) DXA spine imaging is essential in bone health as it enables to evaluate Bone Mineral Density (BMD) and Trabecular Bone Score (TBS), an indirect estimation of trabecular microarchitecture status. Optimal spine segmentation (SpS) is the basis for both BMD and TBS computations to be clinically accurate in regards to diagnosis and individual’s follow-up. Unfortunately, SpS can be a challenging task given the amount and variability of noise in the image, the inter-individual variability in bone shape, rotation, presence of osteophytes, etc. Therefore, SpS in clinical routine often necessitates time-consuming and difficult bone mask manual editing.
The purpose of this study is to implement and optimize a deep learning AI method for an automatic and accurate SpS in AP spine DXA scans.
Method: OsteoLaus is a population-based cohort of 1500 randomly selected Caucasian women (50 to 80 y old) living in Lausanne, Switzerland. We used a subset of 1182 scans acquired on iDXA scanner (GEHC Lunar, Madison, WI, USA) with validated reference bone masks done by bone imaging experts. All scans were anonymized and blind from clinical details. Fifty percent of scans were used for training, 25% for validation, 25% for test. We have adapted the U-net architecture known to be efficient for biomedical image segmentation. Parameters included ReLU activation function, ADAM optimizer for gradient descent and 16 features maps (using Keras with TensorFlow backend). Optimization of the network has been done by minimizing binary cross-entropy loss. Accuracy of the segmentation was evaluated through the Dice coefficient (DC, a similarity index between reference bone mask and network output).
Results: The model rapidly converged and comparison of metrics in both training and validation datasets revealed no model overfitting. Mean (95% CI) DC was found to be 0.981 (0.980 – 0.981), 0.972 (0.965 – 0.979) and 0.960 (0.947 – 0.972) in training, validation and test datasets, respectively. Fig. 1 shows low energy iDXA scans with corresponding reference bone masks and automatic segmentations from the network for 3 random cases of the independent test dataset.
Conclusion: A U-Net based deep learning AI method seems promising for SpS as performance metrics are very good albeit a pilot optimization method was used. Further steps to improve the segmentation will be tested (deeper architecture, specific loss functions, etc.) and should improve model accuracy.