A software program for automated compressive vertebral fracture detection on elderly women’s lateral chest radiograph: Ofeye 1.0

Abstract

Background

Because osteoporotic vertebral fracture (OVF) on chest radiographs is commonly missed in radiological reports, we aimed to develop a software program which offers automated detection of compressive vertebral fracture (CVF) on lateral chest radiographs, and which emphasizes CVF detection specificity with a low false positivity rate.

Methods

For model training, we retrieved 3,991 spine radiograph cases and 1,979 chest radiograph cases from 16 sources, with among them in total 1,404 cases had OVF. For model testing, we retrieved 542 chest radiograph cases and 162 spine radiograph cases from four independent clinics, with among them 215 cases had OVF. All cases were female subjects, and except for 31 training data cases which were spine trauma cases, all the remaining cases were post-menopausal women. Image data included DICOM (Digital Imaging and Communications in Medicine) format, hard film scanned PNG (Portable Network Graphics) format, DICOM exported PNG format, and PACS (Picture Archiving and Communication System) downloaded resolution reduced DICOM format. OVF classification included: minimal and mild grades with <20% or ≥20–25% vertebral height loss respectively, moderate grade with ≥25–40% vertebral height loss, severe grade with ≥40%–2/3 vertebral height loss, and collapsed grade with ≥2/3 vertebral height loss. The CVF detection base model was mainly composed of convolution layers that include convolution kernels of different sizes, pooling layers, up-sampling layers, feature merging layers, and residual modules. When the model loss function could not be further decreased with additional training, the model was considered to be optimal and termed ‘base-model 1.0’. A user-friendly interface was also developed, with the synthesized software termed ‘Ofeye 1.0’.

Results

Counting cases and with minimal and mild OVFs included, base-model 1.0 demonstrated a specificity of 97.1%, a sensitivity of 86%, and an accuracy of 93.9% for the 704 testing cases. In total, 33 OVFs in 30 cases had a false negative reading, which constituted a false negative rate of 14.0% (30/215) by counting all OVF cases. Eighteen OVFs in 15 cases had OVFs of ≥ moderate grades missed, which constituted a false negative rate of 7.0% (15/215, i.e., sensitivity 93%) if only counting cases with ≥ moderate grade OVFs missed. False positive reading was recorded in 13 vertebrae in 13 cases (one vertebra in each case), which constituted a false positivity rate of 2.7% (13/489). These vertebrae with false positivity labeling could be readily differentiated from a true OVF by a human reader. The software Ofeye 1.0 allows ‘batch processing’, for example, 100 radiographs can be processed in a single operation. This software can be integrated into hospital PACS, or installed in a standalone personal computer.

Conclusions

A user-friendly software program was developed for CVF detection on elderly women’s lateral chest radiographs. It has an overall low false positivity rate, and for moderate and severe CVFs an acceptably low false negativity rate. The integration of this software into radiological practice is expected to improve osteoporosis management for elderly women.

Introduction

The bone composition of the spine, which is predominantly trabecular bone, is more prone to the thinning and microarchitectural changes associated with osteoporosis than regions of the hip that are richer in cortical bone. Assessment of vertebral fracture status, in addition to BMD (bone mineral density), provides relevant clinical information to aid in predicting fracture risk in postmenopausal women (1-6). Siris et al. (1) reported that at any given BMD T-score, the risk of incident vertebral, non-vertebral, and any fracture depended heavily on prevalent radiographic osteoporotic vertebral fracture (OVF) status (1). Johansson et al. (2) reported that, in older women and after adjustment for clinical risk factors and BMD, grade-1 OVFs identified on lateral spine imaging with dual-energy x-ray absorptiometry (DXA) are associated with incident major osteoporotic fractures. Though BMD has been commonly used in decision-making for the diagnosis of osteoporosis, it is associated with many inherent limitations (7,8). How to define cutpoint T-scores for differential ethnic groups remains controversial (9). In addition, osteoporotic fracture commonly occurs in subjects not in the category of BMD-defined osteoporosis, and thus would not be selected for pharmacological therapy based on BMD score alone (10-14). On the other hand, OVF is a highly prevalent clinical endpoint (15). OVF can be considered as a “gateway” to other more serious fractures, such as hip fractures.

Many guidelines suggest women with age ≥60 or ≥65 years take osteoporosis screening (16-22). Many fractures and associated complications, including secondary fractures and mortality, can be prevented by routine osteoporosis screening in older people and timely treatment initiation in at-risk individuals. A number of medications have demonstrated that they can reduce fracture risk, both vertebral and nonvertebral, including hip fractures (23-28). Nonpharmacologic approaches to manage osteoporosis, including the combination of weight-bearing and resistance training, adequate calcium and vitamin D intake, and physical activity, can positively affect bone mass. Coupled with preventing falls and limiting modifiable risk factors, such as smoking and alcohol use, these measures can help reduce a person’s risk for osteoporotic fractures (29,30). Early detection of an OVF can lead to further investigation and appropriate management that decreases the risk of future fractures (31). However, osteoporosis screening is still not commonly taken by individuals. In clinical practice, OVF remains under-recognized and undertreated. Because the OVF damage is limited to the anterior vertebral column in most cases, the fracture is usually stable and not associated with neurologic impairment. It is estimated that 3/4 of OVFs are clinically silent. Moreover, OVFs are often unrecognized on chest radiographs, especially when the radiographs were not ordered primarily for skeletal conditions. This underreporting can be due to a lack of awareness by radiologists of the clinical implications of incidental OVFs, and also some radiologists tend to focus their report only on the clinical indication of the X-ray examination, thus tend to not perform an accurate analysis of the spine while reading chest radiograph when the clinical indication is pulmonary or heart diseases (32). Even in back pain patients when the spine is the focus of investigation, OVF can also be missed (33). In our recent retrospective analysis of 105 female cases (mean: 72 years, range: 55–93 years) in a tertiary hospital in China, among the patients with OVF, the false negative reporting rate was 23.9% (17/71, counting cases), though with missed cases being all minimal or mild grades in vertebral height loss. Moreover, in 25 cases with multiple OVFs, in addition to those OVFs reported, one additional OVF was missed in 8 cases, more than one additional OVF were missed in 15 cases, and one additional vertebra with endplate and/or cortex fracture (ECF) was missed in two cases. Multiple and more severe grades of OVFs are associated with an even greater fracture risk (34-36). A precise reporting of the number of OVFs in each patient is also highly relevant in clinical practice.

Because the missing report for OVF on chest radiographs is widespread, computer aided automated detection of compressive vertebral fracture (CVF) (compressive vertebral fracture) is expected to be highly useful (since the diagnosis of OVF should be suggested by a radiologist or physician, computer reading result is termed as CVF, denoting a compressive morphological change of a vertebra). We developed a software program and validated its performance in this study. This version of the software was developed with the following considerations: (I) the software should have the sensitivity of detecting minimal grade CVFs (<20% vertebral height loss); (II) the software should have a low false positivity rate for CVFs; (III) the software should have a low false negativity rate for moderate and severe CVFs (≥25% vertebral height loss). Since we primarily wanted to develop a software program for opportunistic detection of OVF on chest radiographs not indicated for spine disorders, we took the assumption that it will be troublesome to physicians and patients if our software labels a high proportion of false positive reports. On the other hand, it will be acceptable if some minimal and mild OVFs are missed, and it is still acceptable even if a portion of moderate OVFs is missed. Therefore, the goal of our study was to emphasize CVF detection specificity, rather than sensitivity.

Methods

Radiographs materials and OVF gradings

For model training, we retrieved 3,991 spine radiograph cases and 1,979 chest radiograph cases from 15 centers (Table 1). Testing radiographs (542 chest radiograph cases and 162 spine radiograph cases) were from another four independent clinics which differ from the 15 centers which provided training data (Table 1). All cases were female subjects, and except for the training data 12 which were spine trauma cases, all the remaining cases were post-menopausal women. Image data included DICOM format, hard film scanned Portable Network Graphics (PNG) format, DICOM exported PNG format, and Picture Archiving and Communication System (PACS) downloaded reduced resolution DICOM format. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This retrospective study was approved by our institutional ethics committees, study subject consent was obtained for training data 1, 2, 15, and testing data 3, while patient consent was waived for the other data.

Table 1. Information of image data used for ‘base-model’ training and testing.

Hospital	Total cases	Age	OVF cases	X-ray region	Image format
Training data 1	1,994	≥65 years	308	T and L spine	Scanned PNG
Training data 2	1,531	≥69 years	249	T and L spine	Scanned PNG
Training data 3	761	≥65 years	273	Chest	DICOM
Training data 4	127	≥65 years	25	Chest	DICOM
Training data 5	109	≥65 years	32	Chest	DICOM
Training data 6	99	≥65 years	11	Chest	DICOM
Training data 7	100	≥65 years	38	Chest	DICOM
Training data 8	101	≥65 years	36	Chest	DICOM
Training data 9	146	≥65 years	35	Chest	DICOM
Training data 10	284	≥65 years	52	Chest	DICOM
Training data 11	112	≥65 years	27	Chest	DICOM
Training data 12	31	No limitation	31	Spine*	DICOM
Training data 13	132	≥55 years	58	Spine**	DICOM
Training data 14	102	≥65 years	34	Chest	DICOM
Training data 15 #	303	≥60 years	176	T and L spine	Exported PNG
Training data 16	38	≥65 years	19	Chest	DICOM
Total for Training	5,970		1404
Testing data 1	144	≥65 years	49	Chest	DICOM
Testing data 2	164	≥65 years	40	Chest	DICOM
Testing data 3	162	≥85 years	52	T and L spine	Scanned PNG
Testing data 4 ##	234	≥65 years	74	Chest	Reduced DICOM
Total for testing	704		215

Training data 1 and data 2 were acquired for MsOS (Hong Kong) study with 4 years apart, and with different acquisition radiography machines applied, thus equivalent to data from two sources. *, spine traumatic fracture patients (lateral and frontal view radiographs available); **, spine radiograph for back pain patients (lateral and frontal view radiographs available); #, Caucasian subjects from Italy, in DICOM exported PNG format. ##, data in PACS downloaded reduced resolution DICOM format. Only lateral view radiographs available for training data 1, 2, and 15, while all chest radiographs have both lateral and frontal views. OVF, osteoporotic vertebral fracture; scanned PNG, hard film scanned image; DICOM, Digital Imaging and Communications in Medicine; PNG, Portable Network Graphics; PACS, Picture Archiving and Communications System .

The initial readings were based on previous results for spine radiograph OVF epidemiological studies (training data 1, 2, 15, and testing data 3), and chest radiographs were evaluated initially by a trained biomedical engineering graduate (BHX). Reading and labelling were primarily based on lateral radiograph, while frontal radiograph was checked when available and necessary to assist the reading. The final reference readings were established by an experienced radiologist reader (YXJW). The OVF diagnostic criteria were based on previous publications (15,37-43). In addition to vertebral height reductions, attention was paid to alterations in the shape and configuration of a vertebra relative to adjacent vertebrae and expected normal appearances. All the known mimics of OVF were systematically excluded. While a positive ECF sign increases the diagnostic confidence for OVF, an ECF sign was not considered essential for diagnosing an OVF (42). According to the vertebral height loss, OVFs were classified as: minimal and mild grades with <20% or ≥20–25% vertebral height loss respectively, moderate grade with ≥25–40% vertebral height loss, severe grade with ≥40%–2/3 vertebral height loss, and collapsed grade with ≥2/3 vertebral height loss. Since the diagnosis of minimal grade OVF can be sometimes subjective, and these OVFs tend to overall have less immediate clinical relevance, diagnosis of minimal grade OVF was not strictly enforced, i.e., some of the minimal grade OVFs might have been missed during the human reader labelling.

Image pre-processing

Only lateral radiographs were used. Personal information contained in the DICOM format radiographs was removed, and then the images were saved as 1024×1024 PNG format images. Images larger than 1024×1024 were cropped, while images smaller than 1024×1024 were padded with black borders. The contrast and other relevant parameters of the images were adjusted to increase the image clarity.

Based on human labelling, the information of OVF coordinates in the image was obtained.

Measures such as image flipping, rotation at different angles, and random splicing were used to expand the training dataset, for the purpose of improving the robustness and stability of the resulting CVF detection model.

Model training and optimization

Images were normalized to train the CVF detection base model. The detection model structure was composed of three parts, i.e., backbone, neck, and prediction (44), and composed of convolution layers that included convolution kernels of different sizes, pooling layers, up-sampling layers, feature merging layers, and residual modules (45), totaling 213 layers.

The main training parameters were the batch size of 8 (every batch included one negative image), the epoch of 1,000, and the initial learning rate was 0.01. As the number of model iterations increased, the cosine annealing function was used to gradually reduce the learning rate. The dropblock function was used to alleviate the problem of model overfitting. The loss function was (Distance intersection over union) DIOU_Loss, and the (Stochastic Gradient Descent) SGD gradient descent algorithm was used for parameter optimization. Model training was performed using Ubuntu 18.04.4 LTS, python (v3.6.12), pytorch (v1.7.1) machine learning libraries, Nvidia A100(40GB memory) Tensor Core Graphics Processing Unit, and Xenon E5-2698 v4 2.2GHz, 20 Cores Central Processing Unit.

When a CVF is detected, a probability is returned. The model labels a vertebra as with CVF when the probability is ≥0.6. This parameter can be adjusted depending on whether specificity or sensitivity is emphasized.

When the model loss function could not be further decreased, the optimal training model, which we called ‘base-model 1.0’, was obtained.

Base-model 1.0 testing for CVF detection

For testing, in total radiographs of 704 cases were retrieved. These testing radiographs included a mixture of standard DICOM images, film scanned PNG images, and resolution reduced DICOM images downloaded from a PACS. Thus, the testing radiographs were not all of idealized image quality. Lateral radiographs were read by the base-model 1.0, and the reference reading was established by the radiologist reader (YXJW).

After the base-model 1.0 reading, the cases with false positivity and false negativity were further inspected for potential causes.

User interface testing

A user-friendly interface was developed (Figures S1-S4), and the synthesized software was termed Ofeye 1.0. This software allows ‘batch processing’, for example, 100 radiographs can be processed in a single operation. This software can be integrated into hospital PACS, or installed in a standalone personal computer. The clinical usage of a prototype version of the software has been tested externally since Nov 18, 2021, in the Department of Radiology, the First Affiliated Hospital of Zhejiang Chinese Medical University, Hangzhou, China. Users there provided feedback for improving the user-friendliness of this software.

Results

The primary Ofeye 1.0 reading output is demonstrated in Figure 1, the left image window shows the original radiograph; while the right image window shows the radiograph with CVF labelling (if there is/are). If there is no CVF detected, then the left image window and right image window will show the same image. For each suspected CVF detected, a probability is provided. The DICOM view function of this software allows other basic functions including zoom-in, zoom-out, adjusting contrast and window levels, and distance and angle degree measurements.

Figure 1.

The main operation window of Ofeye 1.0. The image on the left is an original lateral chest radiograph. With the image on the right, three CVFs were labelled on this lateral chest radiograph, with a probability of 0.90, 0.91, and 0.83, respectively. Reference reading confirms these three OVFs. One further minimal grade OVF was missed in this case. OVF, osteoporotic vertebral fracture; CVF, compressive vertebral fracture.

CVF detection performance of the base-model 1.0 is shown in Table 2. Testing data 1 and data 2, which were in original DICOM format, showed slightly better performance than testing data 3 (in film scanned PNG format) and testing data 4 (in PACS downloaded reduced resolution DICOM format).

Table 2. Vertebral fracture detection performance of base-model 1.0 (counting cases).

Data	Total	TP case	FP case	TN case	FN case	Sensitivity	Specificity	Accuracy
Data 1	144	43	1	94	6	0.878	0.989	0.951
Data 2	164	34	1	123	6	0.85	0.992	0.957
Data 3	162	46	6	104	6	0.885	0.945	0.926
Data 4	234	62	5	155	12	0.838	0.969	0.927
Total	704	185	13	476	30	0.860	0.973	0.938

TP, true positive; FP, false positive; TN, true negative; FN, false negative.

In total, 33 OVFs in 30 cases had a false negative reading, which constituted a false negative rate of 14.0% (30/215) for counting all OVF cases (including those with minimal or mild grade OVFs). Among these cases with false negative reading, 15 OVFs in 15 cases were mild or minimal grades, while 18 OVFs in 15 cases had OVFs of ≥ moderate grades missed (Table 3), which constituted a false negative rate of 7.0% (15/215) if only counting cases with ≥ moderate grade OVFs. In other words, for cases with ≥ moderate grades OVF, 93% were detected. False positive reading was recorded in 13 vertebrae in 13 cases (one vertebra in each case, Table 4), which constituted a false positivity rate of 2.7% (13/489). These vertebrae with false positivity labeling had a mean probability of 0.72 (ranging from 0.61 to 0.88). They could mostly be easily evaluated by a radiologist reader as being of no significance or a false positive reading (Figure 2).

Table 3. Details of ≥ moderate grade OVFs missed in the independent testing data.

Case number	OVF number	Grade	Explanation for the false negativity
4	4	Moderate	OVF close to a border of the image
1	1	Moderate	OVF overlay with diagram line
1	2	Moderate	OVF together with osteoarthritis
1	3	Moderate	OVF together with osteoarthritis and disappearance of disc space
1	1	Moderate	Oblique filming to the X-ray beam
3	3	Moderate	n/a
2	2	Severe	n/a
1	1	Collapsed	Vertebra totally collapsed
1	1	Collapsed	Vertebra collapsed + rotation

OVFs, osteoporotic vertebral fractures; n/a, not available.

Table 4. Details of false positively labelled vertebrae in the independent testing data.

Case number	OVF number	Explanation for the false positivity
3	3	Labeled vertebra above T4
3	3	Labeled vertebra had rotation to X-ray beam
5	5	Labeled vertebra looks like minimal CVF
1	1	Labeled vertebra looks like minimal CVF + diagram line overlaying
1	1	n/a

OVF, osteoporotic vertebral fracture; CVF, compressive vertebral fracture; n/a, not available.

Figure 2.

A false positive labelling of vertebra T7 in a study subject. The case was from testing data 3 with hard film scanned image. The image was of sub-optimal quality. (A) shows vertebra T7 is labelled as CVF with a relatively low probability of 0.67. (B) is the original image, scoliosis was suspected with multiple thoracic vertebrae showing apparent oval endplate rings. On the other hand, the anterior vertebral height of T7 appears to be similar to the adjacent vertebrae. T7 was not considered to have OVF by reference reading. OVF, osteoporotic vertebral fracture; CVF, compressive vertebral fracture.

Discussion

In recent a few years, a number of authors reported AI (artificial Intelligence) enabled analysis and detection of VF (vertebral fracture) of spine medical images which included spine radiograph (46-53), DXA (54-56), thoracic and/or abdominal CT (57-62), and spine MR images (63-66). Kim et al. (46) presented a structured hierarchical segmentation method that combines the advantages of two deep-learning methods of pose-driven learning and M-net which allows automated detection and segmentation of lumbar vertebrae from radiograph for CVF evaluation. Kim et al. (47) described an approach of AI enabled automated vertebral segmentation of lateral thoracic and lumbar spine radiograph which is expected to be helpful for the measurement of vertebral compression ratio. Seo et al. (48) described a vertebral body segmentation model and a vertebral compression measurement approach on lateral lumbar spine radiographs. Chou et al. (49) and Li et al. (50) reported AI enabled detection of VFs on thoracic and lumbar radiographs, with good accuracy achieved especially for lumbar Genant Grades 2 and 3 VFs. Murata et al. (51) reported AI enabled detection of VFs on plain spinal radiograph. With MRI as the reference standard, Chen et al. (52) reported identifying fresh CVFs from spine radiograph. Chen et al. (53) reported the application of a deep learning algorithm to detect and visualize VFs on plain frontal abdominal radiographs. Derkatch et al. (54) described a model to identify VFs on DXA images. Mehta and Sebro (55) described an application of a support vector machine learning algorithm using posterior-anterior DXA images to identify lumbar spine (L1-L4) VFs without additional lateral DXA imaging. Monchka et al. (56) described an AI model for automated identification of CVF using dual-energy and or single-energy lateral DXA images. Tomita et al. (57) described an AI enabled method to detect incidental CVFs in chest, abdomen, and pelvis CT examinations. Burns et al. (58) described an AI enabled method which detects, localizes, and classifies CVFs and measures BMD of thoracic and lumbar vertebral bodies on CT images. Kolanu et al. (59) described a computer-aided diagnosis model (Zebra Medical Vision) for Genant grades 2 or 3 OVFs in a single center 2357 abdominal and thoracic CT scans. Roux et al. (60) reported automated detection of VF in 150,000 routine lumbar or abdominal CT scans from 35 hospitals in Paris, France. Rueckel et al. (61) described several pathology-specific AI algorithms enabled detection of relevant initially missed secondary thoracic findings in emergency whole-body CT scans, including the detection of cardiomegaly, coronary artery plaques, lung lesions, aortic aneurysms, and VFs. Yabu et al. (63) described AI enabled detection of fresh VF on MR images. Del Lama et al. (64) described AI enabled detection of CVFs on spine MRI images. Yeh et al. (65) described an AI enabled model for the diagnosis of VFs on spine MR images. Yoda et al. (66) described an AI enabled approach for automated differentiation between OVF and CVF due to spinal metastasis on MR images. Overall, most of the reports were single center proof-of-concept studies, with unknown generalisability. Roux et al. (60) reported automated detection of VF in 150,000 routine lumbar or abdominal CT scans from 35 hospitals. However, a shorting coming of the study of Roux et al. appears to be that they did not apply different thresholds (or different criteria) for male and female patients. They reported that, of the patients with VFs, 43.7% were male; while this may represent an over-estimation of VFs in male patients (67).

Compared with literature reports, our base-model 1.0 has the following features: (I) the goal of our software is to detect CVF on lateral chest radiographs (instead of spine radiograph or CT/MRI data) which were initially not indicated for spine disorders. It is less likely that important OVFs are missed by a human reader on spine radiographs taken for spine disorders; (II) our software allows the detection of CVF with <20% height loss, which may be clinically relevant in selected cases (68-71); (III) our software used a large sample size from multiple sources (n=16) for training, the performance of our software was tested with data from multiple external centers (n=4); (IV) despite the testing image data were not all in idealised format, our software offers superior diagnostic performance compared with many literature reports. This base-model 1.0 achieved our goal that (I) it has an overall low false positivity rate (<3% according to our own testing); and (II) it has a low false negativity rate of 7.0% for moderate and severe CVFs (i.e., those of ≥25% vertebral height loss).

There are still many limitations of this study and thus many still limitations for our software Ofeye 1.0. Ofeye 1.0 does not grade the severity of the VF, instead it offers a ‘yes/no’ selection and provides a CVF probability estimation. It labels a vertebra as with CVF when the probability is ≥0.6. Overall, a milder vertebral height loss is associated with a smaller value of the probability and vice versa. In the future, we want to further add the function for CVF grading. Ofeye 1.0 was developed with the goal to emphasize CVF detection specificity at the cost of sensitivity, thus Ofeye 1.0 may not be suitable for traumatic VF assessment. We plan to develop another version of Ofeye which will emphasize CVF detection sensitivity, so that it can be applied to spine traumatic patients. Ofeye 1.0 was trained with radiographs of female subjects, how it can be applied to male subjects requires additional adjustment of the parameters (67), and also additional validations. Ofeye 1.0 was developed for lateral chest radiograph, how it can be reliably applied to the lower lumbar spine also requires additional validations. It should also be noted that, the initial OVF human labelling and the subsequent CVF testing were primarily based on the reference reading provided by one radiologist (YXJW) and his perception of radiographic diagnostics of OVF. However, how to best diagnose and classify minimal and mild grade OVFs remains controversial (72), though there is usually no difficulty for human reader diagnosis of moderate and severe grade OVFs. Thus, it is possible that, if Ofeye 1.0 is tested by a third independent party, slightly different CVF detection performance may be obtained as there is no golden standard for labelling minimal grade OVFs. Finally, it should be noted that Ofeye 1.0 suggests the probability of a vertebra having CVF, it does not offer a firm diagnosis of OVF. The diagnosis of OVF should be made by a radiologist or a physician, by further excluding mimics and other causes (such as artefacts due to poor image quality or scoliosis). While we argue for the importance of recognizing OVFs with <20% vertebral height loss, the clinical management of such OVF cases would depend on the clinical data such as BMD or other fragility fracture history. Though minimal grade OVFs may not have immediate further fragility fracture consequences, they are a biomarker of compromised bone quality. In our MsOS (Hong Kong) year-14 follow-up, out of 150 female participants, five women were identified as having baseline minimal OVF and among them three had osteopenia and two had osteoporosis. There was a trend that these minimal OVF subjects had incident OVF risk similar to that of the subjects with baseline apparent OVF (i.e., ≥20% height loss), higher than female subjects without baseline OVF (6). The real-world importance of minimal/mild OVFs may depend on patients individually, and a wait-and-see strategy with follow-up imaging may be sufficient for many non-traumatic cases. Depending on practical scenarios, for CVFs with minimal/mild extent of vertebral height loss and a small value of probability, the diagnostician may also choose to ignore them for practical reasons even when the CVFs look like being osteoporotic.

In conclusion, we developed a user-friendly software program, Ofeye 1.0, for CVF detection on elderly women’s lateral chest radiographs. This software has an overall low false positivity rate (2.7%), and also for moderate and severe CVFs a low false negativity rate (7.0%). Ofeye 1.0 has batch processing function and can be integrated into hospital PACS. We expect the integration of such a software program into radiological practice will improve osteoporosis management for elderly patients.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This retrospective study was approved by our institutional ethics committees, study subject consent was obtained for training data 1, 2, 15, and testing data 3, while patient consent was waived for the other data.