Automated detection and classification of the proximal humerus fracture by using deep learning algorithm.

Chung SW, Han SS, Lee JW, Oh KS, Kim NR, Yoon JP, Kim JY, Moon SH, Kwon J, Lee HJ, Noh YM, Kim Y. · 2018

Read Automated detection and classification of the proximal humerus fracture by using deep learning algorithm. by Chung SW, Han SS, Lee JW, Oh KS, Kim NR, Yoon JP, Kim JY, Moon SH, Kwon J, Lee HJ, Noh YM, Kim Y. (2018) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc6066766.txt.

Introduction

Chung and co-authors train deep learning to detect and classify shoulder fractures. Their study appeared in Acta Orthopaedica. The team gathered 1,891 plain shoulder X-rays. Specialists labeled normal shoulders and four fracture types. A convolutional neural network learned from the set.

The network spotted fractures with 96 percent top accuracy. Its sensitivity and specificity both neared perfect scores. Sorting the four fracture types proved harder. Accuracy there ranged from 65 to 86 percent. The system beat general doctors and general orthopedists. It matched shoulder specialists overall. Its edge grew on complex three- and four-part breaks. Readers get a strong case for AI help in fracture clinics.

Catalog record

CREATORChung SW, Han SS, Lee JW, Oh KS, Kim NR, Yoon JP, Kim JY, Moon SH, Kwon J, Lee HJ, Noh YM, Kim Y.
YEAR2018
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC6066766

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Automated detection and classification of the proximal humerus fracture by using deep learning algorithm By Seok Won Chung, Seung Seog Han, Ji Whan Lee, Kyung-Soo Oh, Na Ra Kim, Jong Pil Yoon, Joon Yub Kim, Sung Hoon Moon, Jieun Kwon, Hyo-Jin Lee, Young-Min Noh, Youngjun Kim Published in: Acta Orthopaedica DOI: 10.1080/17453674.2018.1453714 ABSTRACT Background and purpose — We aimed to evaluate the ability of artificial intelligence (a deep learning algorithm) to detect and classify proximal humerus fractures using plain anteroposterior shoulder radiographs. Patients and methods — 1,891 images (1 image per person) of normal shoulders (n = 515) and 4 proximal humerus fracture types (greater tuberosity, 346; surgical neck, 514; 3-part, 269; 4-part, 247) classified by 3 specialists were evaluated. We trained a deep convolutional neural network (CNN) after augmentation of a training dataset. The ability of the CNN, as measured by top-1 accuracy, area under receiver operating characteristics curve (AUC), sensitivity/specificity, and Youden index, in comparison with humans (28 general physicians, 11 general orthopedists, and 19 orthopedists specialized in the shoulder) to detect and classify proximal humerus fractures was evaluated. Results — The CNN showed a high performance of 96% top-1 accuracy, 1.00 AUC, 0.99/0.97 sensitivity/specificity, and 0.97 Youden index for distinguishing normal shoulders from proximal humerus fractures. In addition, the CNN showed promising results with 65–86% top-1 accuracy, 0.90–0.98 AUC, 0.88/0.83–0.97/0.94 sensitivity/specificity, and 0.71–0.90 Youden index for classifying fracture type. When compared with the human groups, the CNN showed superior performance to that of general physicians and orthopedists, similar performance to orthopedists specialized in the shoulder, and the superior performance of the CNN was more marked in complex 3- and 4-part fractures. Interpretation — The use of artificial intelligence can accurately detect and classify proximal humerus fractures on plain shoulder AP radiographs. Further studies are necessary to determine the feasibility of applying artificial intelligence in the clinic and whether its use could improve care and outcomes compared with current orthopedic assessments. Proximal humerus fractures are primarily diagnosed using plain radiographs, and the fracture type is determined according to its anatomical location as well as fragmentation and displacement levels. However, since non-orthopedic surgeons or insufficiently experienced orthopedic surgeons are frequently the first doctors to assess fractures, it is not unusual for proximal humerus fractures to be misdiagnosed. In addition, even an experienced orthopedic surgeon can misdiagnose the fracture type due to variable presentation (Mora Guix et al. 2009, Foroohar et al. 2011). Thus, a more efficient and accurate manner of diagnosing and classifying fracture type is of interest. Deep learning is a branch of artificial intelligence that uses a cascade of many layers of nonlinear processing units to extract features and create transformations and is based on the learning of multiple levels of features or representations of the data (Wang and Summers 2012, Bengio et al. 2013, LeCun et al. 2015). Deep learning comprises a neural network with multiple hidden layers that enhance image recognition accuracy, thereby increasing its versatility for capturing representative features (Shin et al. 2013). Since 2012, deep learning has rapidly become the cutting-edge method of enhancing performance in medial image analysis with the use of convolutional neural networks (CNN), which are well suited for analyzing images, and has led to a decrease in the classification error rate from about 25% in 2011 to 3.6% in 2015 (Russakovsky et al. 2015, Lakhani et al. 2017). With such success in identifying and classifying images using a deep learning algorithm, there has been interest in applying deep learning to medical image analysis in several fields, including the detection of skin cancer (Esteva et al. 2017), diabetic retinopathy (Gulshan et al. 2016), mammographic lesions (Kooi et al. 2017), and lung nodules (Hua et al. 2015). However, in the field of orthopedic surgery and traumatology, trials are very scarce despite its importance to public health. To our knowledge, only one study (Olczak et al. 2017) has applied deep learning to fracture orthopedics, and reported promising outcomes of deep learning in identifying fracture, laterality, type of view, and body part. Thus, we aimed to evaluate the diagnostic accuracy of the deep learning algorithm with deep CNN for detecting and classifying proximal humerus fractures using plain anteroposterior (AP) shoulder radiographs. We then compared the results with those of humans. PATIENTS AND METHODS Dataset 1,891 plain shoulder AP radiographs (1,376 proximal humerus fracture cases and 515 normal shoulders) from 1,891 patients (591 men, 1,300 women; 1,083 from Konkuk University Medical Center, 209 from Kyungpook National University Hospital, 165 from Myungji Hospital, 203 from Kangwon National University Hospital, 41 from the National Police Hospital, 25 from Seoul Saint Mary’s Hospital, and 165 from Wonkwang University Sanbon Hospital) were used as the total dataset in this study. We used only 1 image per person to decrease the overperformance of deep learning by the inclusion of a very similar image of the same patient in each test and training set. The mean age of patients was 65 (24–90) years. Fracture classification To evaluate the performance of fracture classification, we classified the proximal humerus fractures into 4 types based on Neer’s classification, which is the most commonly used classification for the proximal humerus fracture: greater tuberosity, surgical neck, 3-part, and 4-part (Neer et al. 1970). A greater tuberosity fracture was defined as 1 displaced fragment of the greater tuberosity component, and A surgical neck fracture as 1 displaced fragment of the surgical neck component. A 3-part fracture was defined as 2 displaced fragments, while a 4-part fracture was defined as having 3 or more displaced fragments from the proximal humerus. In cases of proximal humerus fractures combined with shoulder dislocation (fracture dislocation type), we used the images after reduction and classified them. Each plain shoulder AP radiograph was manually cropped into a square in which the humeral head and neck were centered and constituted approximately 50% of the square image, resized to 256 × 256 pixels, and stored as a JPEG file (Figure 1). [Figure: Each shoulder anteroposterior radiograph was manually cropped into a square in which the humeral head and neck are centered such that they comprise approximately 50% of the square image as illustrated above. Images were then resized to 256 × 256 pixels. Examples of normal and each fracture type: (A) normal, (B) greater tuberosity fracture, (C) surgical neck fracture, (D) 3-part fracture, and (E) 4-part fracture.] Fracture classification was performed by 2 shoulder orthopedic specialists with 14 and 17 years of experience (SWC and KSO) and 1 radiologist with expertise in musculoskeletal diseases and 15 years of experience (NRK). For cases in which the 3 specialists could not agree, the corresponding CT images were checked (CTs were available for all fractures that failed consensus) and then re-discussed. If consensus still could not be achieved even after the evaluation of the CT image(s), the images were excluded from the dataset (n = 21). 346 cases were ultimately classified as greater tuberosity fractures, 514 cases as surgical neck fractures, 269 cases as 3-part fractures, and 247 cases as 4-part fractures. In addition, 515 cases without proximal humerus fractures were classified as the normal group to evaluate the ability of the CNN to distinguish between normal and fractured shoulders (Figure 1). Training of the deep CNN

Read the complete text