TY - CONF
T1 - Automated detection of vertebral fractures in CT using 3D convolutional neural networks
AU - Nicolaes, Joeri
AU - Raeymaeckers, Steven
AU - Robben, David
AU - Wilms, Guido
AU - Vandermeulen, Dirk
AU - Libanati, Cesar
AU - DeBois, Marc
PY - 2020/10/20
Y1 - 2020/10/20
N2 - Objective Summary To develop a fully automated method to identify individual fractured vertebrae in computed tomography (CT) scans. Background• Despite their frequent occurrence and major associated burden, vertebral fractures remain under-diagnosed and patients under-treated. • Spine-containing CT scans provide an opportunity to identify vertebral fractures, yet they commonly go unreported by radiologists. • Automated detection of vertebral fractures would enhance medical care of patients with osteoporosis.Methods• We built a training database of 90 de-identified CT cases, acquired on three different scanners, containing 969 vertebrae scanned for various indications (mean [range] age: 81 [71; 101] years; 64% female; 12% negative cases). • We developed a data-driven, automated vertebral fracture detection method that binarily classifies fractured or normal anatomy for each vertebrapresent in spine-containing CT images.• We performed a stratified five-fold crossvalidation experiment comparing automated predictions with ground truth read-outs from one radiologist.Results• Of all vertebrae in the dataset, 19% were vertebral compression fractures (VCFs) (within expected prevalence for this population).• Dataset contained predominantly lumbar vertebrae and a low number of fractured vertebra between T1–10 (Figure 1).• For our automated predictions compared with the read-outs from a radiologist, the area under the receiver operating characteristic (AUROC)curve was 0.93±0.01 (Figure 2).• Example images of four cases, two with correctly identified VCFs and two with false negative and false positive VCFs, can be found in Figure 3.• These false identifications may have resulted from the limited number of T1–10 fractures in the training set.ConclusionsOur automated vertebral fracture detection method demonstrated the potential for automated early identification of vertebral fractures in patients aged >50 years by opportunistically screening spinecontaining CT images.Confirmatory analysis and additional methodological improvements (e.g. automatic Genant grading, fracture location) using more extensive datasets and method validation are ongoing.
AB - Objective Summary To develop a fully automated method to identify individual fractured vertebrae in computed tomography (CT) scans. Background• Despite their frequent occurrence and major associated burden, vertebral fractures remain under-diagnosed and patients under-treated. • Spine-containing CT scans provide an opportunity to identify vertebral fractures, yet they commonly go unreported by radiologists. • Automated detection of vertebral fractures would enhance medical care of patients with osteoporosis.Methods• We built a training database of 90 de-identified CT cases, acquired on three different scanners, containing 969 vertebrae scanned for various indications (mean [range] age: 81 [71; 101] years; 64% female; 12% negative cases). • We developed a data-driven, automated vertebral fracture detection method that binarily classifies fractured or normal anatomy for each vertebrapresent in spine-containing CT images.• We performed a stratified five-fold crossvalidation experiment comparing automated predictions with ground truth read-outs from one radiologist.Results• Of all vertebrae in the dataset, 19% were vertebral compression fractures (VCFs) (within expected prevalence for this population).• Dataset contained predominantly lumbar vertebrae and a low number of fractured vertebra between T1–10 (Figure 1).• For our automated predictions compared with the read-outs from a radiologist, the area under the receiver operating characteristic (AUROC)curve was 0.93±0.01 (Figure 2).• Example images of four cases, two with correctly identified VCFs and two with false negative and false positive VCFs, can be found in Figure 3.• These false identifications may have resulted from the limited number of T1–10 fractures in the training set.ConclusionsOur automated vertebral fracture detection method demonstrated the potential for automated early identification of vertebral fractures in patients aged >50 years by opportunistically screening spinecontaining CT images.Confirmatory analysis and additional methodological improvements (e.g. automatic Genant grading, fracture location) using more extensive datasets and method validation are ongoing.
M3 - Poster
T2 - European Calcified Tissue Society 2020
Y2 - 21 October 2020 through 24 October 2020
ER -