Keywords:
Artificial Intelligence, Musculoskeletal spine, Digital radiography, CAD, CT, Computer Applications-Detection, diagnosis, Segmentation, Calcifications / Calculi
Authors:
H.-D. Chae, S. H. Hong, H. E. Shin, Y. Lee, M. Park, J. Bae, J.-Y. Choi, H. J. Yoo; Seoul/KR
DOI:
10.26044/ecr2019/C-3569
Methods and materials
We retrospectively collected 1199 lateral cervical radiographs from 444 patients who underwent cervical spine CT between March 2008 and October 2017.
The training set consisted of 699 radiographs from 166 patients diagnosed with OPLL on CT.
The validation data set was composed of 150 radiographs from 38 patients with cervical OPLL and other 150 radiographs from 40 patients without OPLL.
For the test data set,
we used 100 lateral cervical radiographs from 100 patients with cervical OPLL (male-to-female patient ratio,
68:32; mean age 59.4 ± 11.3 years; range 32–81 years) and 100 radiographs from 100 patients without OPLL (male-to-female patient ratio,
54:46; mean age 51.1 ± 16.6 years; range 19–83 years).
The presence or absence of OPLL of the patients in the test data set was confirmed on cervical spine CT,
and the mean interval between CT and radiographs were 13 days for OPLL patients and seven days for the control group.
All the patients with OPLL in the test data set underwent surgical treatment due to cervical myelopathy.
Among them,
74 patients underwent laminoplasty or laminectomy with/without posterior instrumentation,
20 patients underwent anterior cervical discectomy and fusion,
and six patients underwent surgery with both anterior and posterior approach.
The patients in the three data sets were exclusive to each of the other data sets.
A board-certified musculoskeletal radiologist with 4-year of experience in spine imaging placed region-of-interests (ROI) along the boundary of the OPLL on radiographs using commercial software (DeepPhi,
Deepnoid,
Seoul,
Korea).
U-net based neural network was implemented and trained to segment cervical OPLL on lateral radiographs.
DICOM images of lateral cervical radiographs were resized to 480x576 pixels,
and 288x288 sized patches were generated from the image with a stride value of 32.
Among the patches,
patches including at least one pixel of OPLL annotation were used for the training with 200 epochs.
In addition to the quantitative analysis of segmentation accuracy on pixel-level,
the diagnostic performance of the model was evaluated at vertebral- and patient-level,
respectively.