Purpose or Case Report:
Automated segmentation is an important step in automated processing of radiographs. However, manual generation of training data is tedious. To overcome this challenge, we leverage synthetic radiograph training data generated from CT volumes together with corresponding anatomy segmentation labels.
We demonstrate the effectiveness of this approach on the use case of automated scoliosis Cobb angle quantification.
Methods & Materials:
Overview: CT scans and their corresponding vertebral body segmentations were projected to two dimensions to simulate radiographs and radiograph vertebral body labels. These were used as training data for a radiographic vertebral body segmentation model. Cobb angles were then automatically extracted and compared to manual ground truth measurements.
Source Data: 1228 CT scans with 117 different anatomical segmentations of each scan were publicly available , with 913 scans containing vertebral segmentations.
Data Preprocessing, Label Generation & Training: Each CT scan was projected anteroposteriorly generating synthetic radiographs. To augment the data and better represent variable contrast in radiographs, five windowings were generated per subject. Based on these synthetic data, a nnUNet-based deep learning segmentation model was trained to automatically segment vertebrae on radiographs. Based on these segmentations, Cobb angles were automatically calculated based on the slopes of vertebral endplates.
Model Performance: Accuracy of Cobb angle measurements was assessed using data from the MICCAI-AASCE 2019 challenge. Automatically quantified Cobb angles were compared to the reference standard Cobb angles measured by a pediatric radiologist with 15 years experience.
Results:
The model converged, with a mean validation dice coefficient of 0.88. This led to accurate segmentation of both simulated radiographs from CT scans and real radiographs. Qualitative analysis of segmentations on 98 real radiographs revealed only two mild failure cases, each with extreme rotary scoliosis. Predicted Cobb angles had a standard deviation from manually measured Cobb angles of 7.5 and had a mean difference of an average of 5.7 degrees or 26.6 percent when S-shaped scoliosis could correctly be identified.
Conclusions:
Automated labeling of radiographs from CT scans is feasible and facilitates radiographic AI model training. This approach can be extended to various similar tasks, such as leg length quantification.