E2735. Automatic Carpal Bone Segmentation on Plain Radiographs Using Deep-Learning Network Trained With Digitally Reconstructed Radiographs From CT
Authors
Hee-Dong Chae;
Seoul National University Hospital
Sung Hwan Hong;
Seoul National University Hospital
Ja-Young Choi;
Seoul National University Hospital
Hye Jin Yoo;
Seoul National University Hospital
Ji Sun Oh;
Seoul National University Hospital
Objective:
Accurate bone segmentation is necessary to automatically measure many orthopedic parameters. However, it is often difficult to determine the exact boundary of bone structures on a 2D projection image such as a plain radiograph due to the overlap of bone with other tissues. The digitally reconstructed radiography (DRR) is a synthetic x-ray generated from the CT by projecting the 3D volumetric data onto the 2D plane. We trained a deep-learning model with DRRs to automatically segment carpal bones on the actual wrist anteroposterior (AP) and lateral radiographs.
Materials and Methods:
Two hundred wrist CT scans were used to generate DRRs, and 15 bone segmentation masks (distal radius, distal ulna, five metacarpal bones, and eight carpal bones) were also generated for each DRR using the CT segmentation mask. DRR dataset were split into training set (n = 100), validation set (n = 50) and test set (n = 50). A pair of DRRs composed of wrist AP and lateral images were generated from one CT, and data augmentation was performed by changing the projection angle when generating the train set. DeepLab v3+ with EfficientNet-B0 encoder was used to develop the segmentation model. The performance of the segmentation model was evaluated in the test set composed of DRR and the test set composed of real wrist AP and lateral radiograph, respectively. For the real radiograph test set, we used 50 wrist AP and lateral with manually segmented ground truth masks. The accuracy of the segmentation model was evaluated using the Dice similarity coefficient (DSC).
Results:
On DRRs, the deep-learning model showed good performance for the segmentation of wrist bones (mean DSC range, 0.67-0.95). In distal radius and distal ulna, mean DSC were 0.95 and 0.94 in wrist AP, and 0.93 and 0.91 in wrist lateral, respectively. The mean DSC of 8 carpal bones in wrist AP DRR was 0.79-0.93, and 0.67-0.88 in lateral DRR. Among carpal bones, the segmentation accuracy in the scaphoid was the highest, and the performance in the hamate was the lowest. On real wrist radiographs, the segmentation model showed moderate to good accuracy and was slightly lower than that of DRR (mean DSC range, 0.56-0.91). In distal radius and distal ulna, mean DSC were 0.89 and 0.91 in wrist AP, and 0.87 and 0.83 in wrist lateral, respectively. The mean DSC of 8 carpal bones in real wrist AP was 0.56-0.85, and 0.59-0.86 in lateral radiographs.
Conclusion:
We developed and validated a deep-learning based bone segmentation model on wrist AP and lateral radiographs using DRRs. DRR-based segmentation model can segment each bone relatively accurately, even in lateral radiographs where bone structures are largely overlapped.
