Automated Adolescence Scoliosis Detection Using Augmented U-Net With Non-square Kernels

Abstract

French

Purpose: Scoliosis is a deformity of the spine, and as a measure of scoliosis severity, Cobb angle is fundamental to the diagnosis of deformities that require treatment. Conventional Cobb angle measurement and assessment is usually done manually, which is inherently time-consuming, and associated with high inter- and intra-observer variability. While there exist automatic scoliosis measurement methods, they suffer from insufficient accuracy. In this work, we propose a two-step segmentation-based deep learning architecture to automate Cobb angle measurement for scoliosis assessment using X-Ray images. Methods: The proposed architecture involves two steps. In the first step, we utilize a novel Augmented U-Net architecture to generate segmentations of vertebrae. The second step includes a non-learning-based pipeline to extract landmark coordinates from the segmented vertebrae and filter undesirable landmarks. Results: Our proposed Augmented U-Net architecture achieved a Symmetric Mean Absolute Percentage Error of 9.2%, with approximately 90% of estimations having less than 10 degrees difference compared with the AASCE-MICCAI challenge 2019 dataset ground truths. We further validated the model using an internal dataset and achieved almost the same level of performance. Conclusion: The proposed architecture is robust in providing automated spinal vertebrae segmentations and Cobb angle measurement, and is potentially generalizable to real-world clinical settings.

Visual Abstract

Keywords

Cobb angle vertebra landmark scoliosis convolutional neural network

Introduction

Scoliosis refers to an abnormal curvature of the spine, with signs such as back pain and shoulder and waist asymmetry.¹ As the reference for scoliosis diagnosis and treatment planning,² Cobb angles in X-Ray images measure the curvature of the spine. Lines parallel to the end vertebrae in the spine curve form the basis for Cobb angle calculation.³ Two X-Ray examples of curved spines are shown in Figure 1. Positive judgement is made on the presence of scoliosis if the largest Cobb angle exceeds 10 degrees, and the severity rises when it reaches 25 degrees.¹ The conventional standard manual approach for measuring the Cobb angle is associated with high inter- and intra-observer errors due to the variation in the level of operator’s experience and the nature of manual measurement.⁴ Automated Cobb angle estimation techniques have been proposed to alleviate the problems caused by manual measurements.

Figure 1.

(a) Key vertebrae identifications for Cobb angle measurement on frontal view spine radiograph (marked in blue) with curved spine from AASCE-MICCAI challenge 2019 dataset.⁵ (b) Spine subregions for Cobb angle measurement on frontal view spine radiograph (marked in red) with curved spine from the internal dataset.

Previous Work

The majority of the current computer-aided automatic Cobb angle estimation techniques are based on the task of localizing landmarks in the thoracic and lumbar regions. These pipelines use landmark coordinates to compute Cobb angles. Common approaches to localizing landmarks include direct estimation and segmentation of the vertebras.

Direct estimation methods

Direct estimation approaches for the vertebra landmarks approximation have adopted various machine learning (ML) algorithms, such as Support Vector Regression (SVR) and Convolutional Neural Networks (CNN). Sun et al proposed the Structured Support Vector Regression (S²VR)⁶ technique, which takes the advantage of SVR in handling non-linear input-output relationships, and relies on mapping hand-crafted features from input X-Ray images to landmark coordinates, and Cobb angles simultaneously. BoostNet⁵ is another direct estimation approach that automates feature extraction using a CNN and filters outlier features with statistical methods. Khanal et al. proposed a CNN-based solution⁷ to isolate each vertebra with Faster-RCNN,⁸ which consists of a Region Proposal Network (RPN) to propose candidate regions of interest (ROIs) and a layer to regress vertebrae bounding boxes for each individual vertebra. Once the bounding boxes are formed, landmarks within each individual vertebra bounding box are detected by a DenseNet.⁹

Segmentation-Based Methods

Segmentation-based approaches for the vertebra landmarks approximation consist of classifying individual pixels into foreground and background, followed by extracting landmark coordinates from the segmentations. Horng et al. utilized CNNs for vertebrae segmentation¹⁰ and used three variations of U-Net architecture,¹¹ namely, the base U-Net, Residual U-Net, and Dense U-Net. Residual U-Net achieved the best results. Anitha et al. proposed Active Contour method,¹² where the vertebrae are contoured by gradient vector flow (GVF) snakes.¹³ GVF snakes are initialized with a closed path encompassing the ROI which is iteratively improved based on the integration of gradients of pixel intensities on the path.

The accuracy of landmark localization techniques is crucial for predicting Cobb Angles. Direct estimation approaches are impacted by the highly non-linear association between input images and desired output, and thus any minor error in identifying landmark coordinates may lead to significant errors in Cobb angle calculations. In contrast, segmentation-based approaches extract vertebra landmarks from processed segmentations. Nevertheless, X-Ray machine settings, anatomical variability, and low tissue contrast can affect segmentation results. In this paper, we propose a novel architecture to tackle the challenges of the segmentation-based methods by introducing extra kernel constraints when selecting image features.

Methodology

Research Ethics Board approval was obtained for this study. Given the retrospective nature of the study informed consent was waived.

Dataset

AASCE-MICCAI challenge 2019 dataset

AASCE-MICCAI challenge 2019 dataset⁵ is composed of 609 spinal Anterior-Posterior (AP) X-Ray images, with all of the patients showing signs of scoliosis. Since the dataset includes cervical vertebrae, which are seldom involved in spinal deformity,¹⁴ all images were cropped to only include the ROIs (i.e., 17 vertebrae from the thoracic and lumbar regions). Each image is provided with 68 landmark coordinates labelled by professional radiologists. The landmark ground truths are ordered from top to bottom, left to right, and normalized into [0, 1]. They also include 3 Cobb angles in the order of proximal thoracic (PT), main thoracic (MT), and thoraco-lumbar (TL) spine. We randomly split the dataset into subsets of 481 samples for training (85%) and validation (15%) and 128 samples for testing. In the test set, 26 cases were excluded due to inaccurate ground truth landmarks, leaving the test set with 102 samples. Figure 1(a) shows an example from the AASCE-MICCAI challenge 2019 dataset.⁵

Internal dataset

We further validated (external validation) the proposed architecture on another dataset obtained internally in a paediatric hospital. The dataset is composed of 55 spinal standing AP X-Ray images captured from children aged 9 to 17 showing signs of scoliosis. A subset of patients (n = 8) who had undergone prior spine surgeries, who were using braces during the X-Ray data acquisition, or who had spine segmentation anomalies were marked as extreme cases and hence, two datasets were constructed for the external validation: the first set that includes all cases (n = 55), and the second set that excludes extreme cases (n = 47).

The images in the internal dataset were manually cropped to reproduce AASCE-MICCAI challenge 2019 dataset’s constraints on the ROIs. The internal dataset contains similar ground truth information as the AASCE-MICCAI challenge 2019 dataset. The landmark coordinates were labelled by a senior radiology resident (PGY5) (S.H.) who was supervised by a senior radiology staff with more than 20 years of experience in pediatric radiology (A.S.D.). It should be highlighted that the internal dataset was not involved in model training, and all samples in the dataset were reserved for testing purposes. Figure 1(b) shows an example of the internal dataset.

Data augmentation

Given CNNs demanding large training datasets, data augmentation is often used in pipelines.¹⁵ By augmenting the dataset, the CNN is exposed to more diversity in the data, and thus is less likely to overfit the dataset.¹⁵ We employed conventional data augmentation techniques including mirroring each data sample, randomly tilting each image with an angle up to 10 degrees, randomly adjusting the gamma of each image, and the combination of multiple operations, for example, tilting after the gamma is adjusted.

Proposed Architecture

The proposed architecture consists of two phases: (1) Segmenting objects of interest by an Augmented U-Net neural network and (2) refining extracted landmark coordinates by a non-learning-based error correction and outlier rejection technique.

The augmented U-net architecture

The proposed Augmented U-Net is a novel architecture to perform precise segmentation tasks on spinal vertebra region. The architecture is a derivation of the conventional U-Net architecture,¹¹ with an additional path, namely, the Augmented path. The Augmented U-Net is trained to segment the vertebrae, where the target shape is rectangular.¹⁶ Literature suggests the average length of the vertebrae is approximately twice their average height,^16–18 which was confirmed by the experiments we conducted on a sample dataset from the training cohort. Hence, we enforced the model to treat the horizontal pixels twice as significant compared with vertical pixels by applying non-square kernels on the augmented path. Inspired by the Double U-Net,¹⁹ we hypothesized the output feature map produced by using only non-square kernels can be further refined by incorporating feature map produced by square kernels. Consequently, a path containing square kernels, which simulated conventional U-Net architecture¹¹ was employed in parallel. This ensures that the Augmented U-Net does not lose the information flow of the original U-Net architecture while enforcing a priori knowledge of the shape of the vertebrae. The Augmented U-Net learnt features from both settings, and two paths were merged before final output generation.

Non-square kernels

We enforced vertebra shape constraints by introducing non-square kernels. The non-square kernels assign more attention to the horizontal features than the vertical features within the filters. Nevertheless, non-square kernels may fade edge and corner features in the feature maps because they have a broader field of view. Thus, a conventional U-Net branch is employed in parallel to the non-square-kernel path.

Implementation

A block diagram of the proposed Augmented U-Net architecture is shown in Figure 2. Each of the paths in the Augmented U-Net architecture behaves as an individual U-Net network in the early stage. The input images were cropped, and their size was unified to 256 × 512, resulting in ROI-limited frames. The input images were fed into both paths simultaneously. Each path started with dedicated encoder structures to extract image features. The encoder structures consist of multiple convolutional layers and max-pooling layers. Max-pooling layers downsample the input images by a multiplier of 4 after every two convolutional layers. Dropout layers were then employed in between consecutive convolutional layers to prevent model from overfitting.

Figure 2.

Block diagram of the proposed Augmented U-Net architecture.

Decoders were then applied to project lower-level features back to the pixel space. The decoders consist of multiple convolutional layers and upsampling layers. Similar to the encoder, we used upsampling layers between every two convolutional layers and dropout layers between consecutive convolutional layers. Similar to the original U-Net architecture,¹¹ the Augmented U-Net requires concatenations between encoders and decoders to refine details that may vanish during downsampling.¹¹ The two-channel output layer of the Augmented U-Net was activated by the Softmax function, corresponding to the probability of each pixel belonging to the object of interest (foreground) and background, respectively. We compared both probabilities and classified each pixel to either 1 (foreground) or 0 (background) accordingly.

Vertebra Extraction

We applied Border-following algorithm²⁰ to the segmentation results to initiate the vertebra extraction process. This algorithm detects the borders of connected foreground objects in the segmentation results, namely, the vertebrae, and reformats them into collections of coordinates. Within each vertebra contour, the minimum bounding rectangle (MBR) algorithm²¹ was used to discover the tightest enclosing rectangle.

Outlier Rejection

The following sections discuss error correction and outlier rejection techniques we introduced to refine the extracted landmark coordinates based on the adjacent vertebrae.

Location filtration

Spine vertebrae form a line when they are connected, and consecutive vertebrae have close orientations and coordinates. A segmentation generated by the Augmented U-Net may contain misclassified pixels in the form of outliers. We used the coordinates of the topmost vertebra, which is assumed to be correctly segmented, as a reference to enforce the location offsets of consecutive vertebrae. We determined the maximum offset to be 40 pixels horizontally based on experiments on the training set. Vertebrae farther from their adjacent neighbours were treated as outliers and thus, discarded.

Size filtration

The Augmented U-Net network occasionally generates false positive objects in the spine area other than vertebrae, such as tissues and ribs. We studied the size variation between true positives and false positives from ground-truth information in the training cohort. The results suggested the average size of the vertebrae above and below the apex were similar. Hence, we grouped foreground objects based on their positions related to the apex and calculated the average sizes for both sets. Based on the experiments, the foreground objects with a size 60% larger than the average size were considered non-vertebra objects and thus, discarded.

Cobb Angle Calculation

We extracted the landmarks of the four corners of each vertebra and derived the final Cobb angles as follows: (1) compute slopes for every pair (left and right corners) of vertebrae landmarks, (2) identify the landmark pair with the furthest deviation from the expected position of the spine, that is, the apex,²² (3) identify the most tilted landmark pair (i.e., with the largest slope) above the apex (α), (4) identify the most tilted landmark pair (i.e., with the largest slope) below the apex (β), and (5) compute 3 Cobb angles using equation (1):

a) proximal thoracic (PT) cobb angle – angle between the topmost landmark pair and α.

b) main thoracic (MT) cobb angle – angle between α and β.

c) thoraco-lumbar (TL) cobb angle – angle between β and the bottom most pair

Arctan θ = \frac{s l o p e_{1} - s l o p e_{2}}{1 + s l o p e_{1} \times s l o p e_{2}}

(1)

where θ is each of the three Cobb angles,

s l o p e_{1}

and

s l o p e_{2}

are the slopes of the two landmark pairs. Figure 1. (a) indicates key vertebrae identifications for this process.

Evaluation Metrics

We used the Symmetric Mean Absolute Percentage Error or SMAPE, as our primary evaluation metric to quantify the quality of the predicted Cobb Angles. Equation (2) describes the SMAPE metric, where N represents the number of samples in the test dataset, and A and B are the ground truth and predicted angles, respectively. Lower SMAPE corresponds to a better performance

S M A P E = \frac{1}{N} \sum_{N} \frac{S U M [| A - B |]}{S U M [A + B]} \times 100 %

(2)

The actual angle differences calculated according to equation (3) are also important to compare the predictions and ground truths

∆ A n g l e = | A - B |

(3)

Results

The novel architecture proposed in this paper achieved superior performances for both spinal vertebra segmentation and Cobb angle measurement tasks compared to benchmarks.^10,23,24

Qualitative Results for Spinal Vertebra Segmentation

Figure 3 illustrates the qualitative segmentation results of spinal vertebra generated by the Augmented U-Net architecture as highlighted areas. At least 17 segmented vertebral endplates were included in each image. Exceptions existed as the X-Ray images may include vertebral endplates other than thoracic and lumbar vertebra. The proposed architecture accurately detected all desired spinal vertebrae (highlighted in red) and generated precise segmentations. Unlabelled or partially labelled vertebrae, for example, the bottom vertebra from Figure 3(a), were discarded by our proposed outlier rejection algorithm. The segmented vertebrae followed the curve of the spine, and clear boundaries between adjacent vertebrae were presented in the majority of cases (87 out of 102). Successful separation of the adjacent vertebrae improved the precision of the later landmark coordinates extraction step.

Figure 3.

Qualitative results for spinal vertebra segmentation on (a) an X-Ray image and (b) an X-Ray image with low background contrast ratio. The red points highlight estimated landmark coordinates.

Quantitative Results for Cobb Angles Measurement

Our proposed architecture achieved an overall SMAPE of 9.2% for the entire test set in the AASCE-MICCAI challenge 2019 dataset,⁵ with SMAPE of PT, MT, and TL being 5.27%, 9.27%, and 12.48%, respectively. The robustness of the architecture was further confirmed by the absolute angle differences between predictions and ground truths, as shown in Table 1. Over 92% of predictions had less than 10 degrees of absolute differences from the corresponding ground truths, with the majority being less than 5 degrees.

Table 1.

Performance of the Proposed Augmented U-Net Architecture on the AASCE-MICCAI Challenge 2019 Dataset⁵; Over 92% of the Predicted Cobb Angles had an Error Less than 10 degrees.

Absolute angle difference	% of cases
Absolute angle difference ≤ 5°	72.21
5° <Absolute angle differences ≤ 10°	20.59
10° <Absolute angle differences ≤ 20°	4.58
Absolute angle differences > 20°	2.62

External Validation

We further validated our proposed architecture on the internal dataset. Table 2 shows that the architecture generalized well to the internal dataset; over 90% of the predicted Cobb angles in the internal dataset (including the extreme cases) had an error less than 10 degrees. The overall SMAPE of 12.9% for the internal dataset (including the extreme cases) was close to the performance on the AASCE-MICCAI challenge 2019 dataset⁵ (SMAPE: 9.2%).

Table 2.

Performance of the Proposed Augmented U-Net Architecture on the Internal dataset With and Without Extreme Cases; Over 90% of the Predicted Cobb Angles in Both Cases had an Error Less than 10 degrees.

	Including extreme cases	Excluding extreme cases
SMAPE	12.9%	12.6%
SMAPE	% of cases	% of cases
Absolute angle difference ≤ 5°	68.50%	71.60%
5° <Absolute angle differences ≤ 10°	21.80%	22.00%
10° <Absolute angle differences ≤ 20°	7.30%	5.70%
Absolute angle differences >20°	2.40%	0.70%

Figure 4 illustrates examples of extreme cases in the internal dataset, where large artifacts and pixel noise are observed around the vertebra region. Such artifacts resulted in reduced visual distinguishability of nearby vertebrae. We investigated the performance of the architecture after excluding the extreme cases that the radiologists could not confidently distinguish vertebra landmarks due to the interfering noises. The overall SMAPE without extreme cases was 12.6%, and over 93% of Cobb angle predictions had absolute differences of fewer than 10 degrees from the ground truths.

Figure 4.

Examples of extreme cases in the internal dataset.

Vertebrae Segmentation Effect

We compared the performance of conventional U-Net, U-Net with non-square kernels, and the proposed Augmented U-Net architecture on the AASCE-MICCAI challenge 2019 dataset,⁵ while the network architectures, input size, and hyperparameters were kept similar. As shown in Table 3, the segmentation generated by U-Net with non-square kernel yielded the most inaccurate Cobb Angle measurement, with a SMAPE of 15.7%. The conventional U-Net¹¹ improved the SMAPE to 11.6%, and the proposed Augmented U-Net further refined the vertebra segmentation resulting in a SMAPE of 9.2%. Table 3 also includes the proportions of predictions in different absolute angle difference bins, showing over 72% of the Augmented U-Net predictions have an absolute angle difference less than 5% compared to the ground truths.

Table 3.

Performance of the Proposed Augmented U-Net Architecture on the AASCE-MICCAI Challenge 2019 Dataset⁵ Compared to the Conventional U-Net, and U-Net With Non-Square Kernel.

	U-Net	U-Net with non-square kernel	Augmented U-Net
SMAPE	11.6%	15.7%	9.2%
SMAPE	% of cases	% of cases	% of cases
Absolute angle differences ≤ 5°	59.15%	53.59%	72.21%
5° <Absolute angle differences ≤ 10°	28.76%	24.18%	20.59%
10°<Absolute angle differences ≤ 20°	9.15%	15.69%	4.58%
Absolute angle differences > 20°	2.94%	6.54%	2.62%

Comparison to the State-of-the-Art

We compared the performance of our model on the AASCE-MICCAI challenge 2019 dataset⁵ with the state-of-the-art, including two regression approaches (ResNet,²³ Extrapolation Net²⁴ and Faster-RCNN + DenseNet⁷) and one CNN-based segmentation method.¹⁰ Table 4 shows the SMAPE results of each method. In comparison to the benchmarks, our architecture achieved the lowest SMAPE of 9.2%, which was significantly lower than the previous segmentation-based benchmark¹⁰ (SMAPE of 16.5%).

Table 4.

Comparison of SMAPE Results of the Proposed Augmented U-Net Architecture With Benchmarks Using the AASCE-MICCAI Challenge 2019 Dataset.⁵

Method	SMAPE (%)
ResNet (regression)²³	10.8
Extrapolation net (regression)²⁴	23.4
Faster-RCNN + DenseNet (regression)⁷	25.7
ConvNets (segmentation)¹⁰	16.5
Augmented U-Net (segmentation)	9.2

Discussion

In this work, we proposed Augmented U-Net segmentation architecture, which is robust in predicting Cobb angles from AP X-Ray images. The proposed architecture enhances vertebrae segmentation by applying non-square kernels, resulting in clear boundaries between adjacent vertebrae and fewer outliers. The post-processing algorithms efficiently filter undesired foreground pixels.

The common approach for automatically computing Cobb angles is based on localizing landmark coordinates of vertebrae in X-Rays, which is usually done by either direct estimation methods such as SVR⁶ and BoostNet⁵ or segmentation-based methods, which are usually based on variations of U-Net architecture.¹¹ Direct estimation methods are sensitive to accurate landmark localization, and conventional U-Net based segmentation methods do not consider the rectangle shape of vertebrae. Our proposed architecture overcame these limitations, and outperformed the existing methods (SMAPE: 9.2% vs 10.8%-25.7%). The external validation of our proposed architecture further confirmed the robustness of the Augmented U-Net in the estimation of Cobb angles (SMAPE: 12.9%).

The level of precision achieved by our proposed architecture enables radiologists in real-world clinical settings to perform faster, more accurate, and more reliable diagnostics for scoliosis.²⁵ Moreover, the vertebra segmentations generated by the Augmented U-Net can be used for purposes such as spinal surgery planning.

This work can be improved if a larger, multi-institutional dataset is accessible. Another clinically important point is separating patients who had surgery for training and testing. Larger batches of data for training and a deeper architecture could help the performance, however, training the Augmented U-Net would become computationally expensive.

Conclusion

In this paper, we proposed a novel segmentation architecture for spinal Cobb angle estimation, which takes advantage of the prior knowledge of the target shapes and applies designated non-square kernels to acquire superior segmentation results. Our approach resulted in improved SMAPE and absolute angle differences on the AASCE-MICCAI challenge 2019 dataset compared to benchmarks, demonstrating the pipeline’s applicability in clinical settings. Our architecture also achieved similar performance on an external validation.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Chair in Medical Imaging and Artificial Intelligence, a joint Hospital-University Chair between the University of Toronto, The Hospital for Sick Children, and the SickKids Foundation.

ORCID iDs

Khashayar Namdar

Farzad Khalvati

Appendix

The pretrained Augmented U-Net model can be accessed via: https://github.com/IMICSLab/Augmented_U-Net_for_Scoliosis_Detection.git

References

Janicki

Alman

. Scoliosis: Review of diagnosis and treatment. Paediatr Child Health. 2007;12(9):771-776. doi:10.1093/pch/12.9.771

Langensiepen

Semler

Sobottke

, et al. Measuring procedures to determine the Cobb angle in idiopathic scoliosis: A systematic review. Eur Spine J. 2013;22(11):2360-2371. doi:10.1007/s00586-013-2693-9

Tanure

Pinheiro

Oliveira

. Reliability assessment of Cobb angle measurements using manual and digital methods. Spine J. 2010;10(9):769-774. doi:10.1016/j.spinee.2010.02.020

Gstoettner

Sekyra

Walochnik

Winter

Wachter

Bach

. Inter- and intraobserver reliability assessment of the Cobb angle: Manual versus digital measurement tools. Eur Spine J. 2007;16(10):1587-1592. doi:10.1007/s00586-007-0401-3

Bailey

Rasoulinejad

. Automatic landmark estimation for adolescent idiopathic scoliosis assessment using BoostNet. In: Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) . Vol 10433 LNCS; 2017. doi:10.1007/978-3-319-66182-7_15

Sun

Zhen

Bailey

Rasoulinejad

Yin

. Direct estimation of spinal cobb angles by structured multi-output regression. In: Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) . Vol 10265 LNCS; 2017. doi:10.1007/978-3-319-59050-9_42

Khanal

Dahal

Adhikari

Khanal

. Automatic cobb angle detection using vertebra detector and vertebra corners regression. In: Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) . Vol 11963 LNCS; 2020. doi:10.1007/978-3-030-39752-4_9

Ren

Girshick

Sun

. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE Trans Pattern Anal Mach Intell. 2017;39(6):1137-1149. doi:10.1109/TPAMI.2016.2577031

Huang

Liu

van der Maaten

Weinberger

. Densely connected convolutional networks. In: Proceedings - 30th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017 , January, 2017. doi:10.1109/CVPR.2017.243

10.

Horng

Kuok

Lin

Sun

. Cobb angle measurement of spine from x-ray images using convolutional neural network. Comput Math Methods Med. 2019;2019:6357171. doi:10.1155/2019/6357171

11.

Ronneberger

Fischer

Brox

. U-net: Convolutional networks for biomedical image segmentation. In: Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) . Vol 9351; 2015. doi:10.1007/978-3-319-24574-4_28

12.

Anitha

Prabhu

. Automatic quantification of spinal curvature in scoliotic radiograph using image processing. J Med Syst. 2012;36(3):1943-1951. doi:10.1007/s10916-011-9654-9

13.

Prince

. Snakes, shapes, and gradient vector flow. IEEE Trans Image Process. 1998;7(3):359-369. doi:10.1109/83.661186

14.

Diebo

Shah

Nv.

Boachie-Adjei

, et al. Adult spinal deformity. Lancet. 2019;394(10193):160-172. doi:10.1016/S0140-6736(19)31125-0

15.

Shorten

Khoshgoftaar

. A survey on image data augmentation for deep learning. J Big Data. 2019;6(1):60. doi:10.1186/s40537-019-0197-0

16.

Masharawi

Salame

Mirovsky

, et al. Vertebral body shape variation in the thoracic and lumbar spine: Characterization of its asymmetry and wedging. Clin Anat. 2008;21(1):46-54. doi:10.1002/ca.20532

17.

Zhou

McCarthy

McGregor

Coombs

RRH

Hughes

SPF

. Geometrical dimensions of the lower lumbar vertebrae--analysis of data from digitised CT images. Eur Spine J. 2000;9(3):242-248. doi:10.1007/s005860000140

18.

Lee

Kwak

, et al. Determination of sex for the 12th thoracic vertebra by morphometry of three-dimensional reconstructed vertebral models. J Forensic Sci. 2008;53(3):620-625. doi:10.1111/j.1556-4029.2008.00701.x

19.

Jha

Riegler

Johansen

Halvorsen

Johansen

. DoubleU-Net: A deep convolutional neural network for medical image segmentation. In: Proceedings - IEEE Symposium on Computer-Based Medical Systems. 2020-July. doi:10.1109/CBMS49503.2020.00111

20.

Suzuki

. Topological structural analysis of digitized binary images by border following. Comput Vis Graph Image Process. 1985;30(1):32-46. doi:10.1016/0734-189X(85)90016-7

21.

Chaudhuri

Samal

. A simple method for fitting of bounding rectangle to closed regions. Pattern Recognit DAGM. 2007;40(7):1981-1989. doi:10.1016/j.patcog.2006.08.003

22.

Erratum to: Scoliosis imaging: What radiologists should know (RadioGraphics, 30, (7), (1823–1842). doi:10.1148/rg.307105061). Radiographics. 2015;35(4).

23.

Huang

Metaxas

. Vertebra-focused landmark detection for scoliosis assessment. In: Proceedings - International Symposium on Biomedical Imaging. April, 2020. doi:10.1109/ISBI45749.2020.9098675

24.

Wang

Leung

Chung

Chen

. Accurate automated Cobb angles estimation using multi-view extrapolation net. Med Image Anal. 2019;58:101542. doi:10.1016/j.media.2019.101542

25.

Morrissy

Goldsmith

Hall

Kehl

Cowie

. Measurement of the Cobb angle on radiographs of patients who have scoliosis. Evaluation of intrinsic error. J Bone Joint Surg Am. 1990;72(3):320-327. doi:10.2106/00004623-199072030-00002