Demir Sukru, Key Sefa, Tuncer Turker, Dogan Sengul
Department of Orthopedic, Firat University Faculty of Medicine, Elazığ, Turkey.
Department of Orthopedic, Bingöl State Hospital, Bingöl, Turkey.
Med Hypotheses. 2020 Mar 4;140:109663. doi: 10.1016/j.mehy.2020.109663.
Humerus fracture have been widely seen disease in the orthopedic clinics and classification of them is a hard process for orthopedist. The main aim of the proposed method is to classify humerus fracture by using a naïve and multileveled method. We collected a novel humerus fracture X-ray image dataset. This dataset consists of 115 images. In this paper, a novel stable feature extraction method is presented to classify humerus fractures. This method is called exemplar pyramid method and it is inspired by exemplar facial expression recognition methods. To classify humerus fractures, X-ray images were employed as input. In this study, X-ray images are resized to 512 × 512 sized image. Then, the used humerus fracture images are divided into 64 × 64 size of exemplars. To create levels, maximum pooling which has been mostly used in deep networks is used and four levels are created. Histogram of oriented gradients (HOG) and local binary pattern (LBP) are employed for feature generation. The most discriminative ones of the generated and concatenated features are selected by using ReliefF and Neighborhood Component Analysis (NCA) based two levelled feature selector (RFNCA). To emphasize success of the proposed exemplar pyramid model based feature generation, four conventional classifiers are chosen for classification and the proposed exemplar pyramid model achieved 99.12% classification accuracy by using leave one out cross validation (LOOCV). Results and tests clearly illustrates success of the proposed exemplar pyramid model based humerus fracture classification method. The results also shown that the proposed exemplar pyramid model achieved higher classification rate than Orthopedist specialized in shoulder.
肱骨骨折是骨科门诊中常见的疾病，对其进行分类对骨科医生来说是一个困难的过程。所提出方法的主要目的是使用一种简单且多层次的方法对肱骨骨折进行分类。我们收集了一个新颖的肱骨骨折X线图像数据集。该数据集由115张图像组成。本文提出了一种新颖的稳定特征提取方法来对肱骨骨折进行分类。这种方法被称为示例金字塔方法，它受到示例面部表情识别方法的启发。为了对肱骨骨折进行分类，将X线图像用作输入。在本研究中，将X线图像调整为512×512大小的图像。然后，将使用的肱骨骨折图像划分为64×64大小的示例。为了创建层次，使用了深度网络中最常用的最大池化，并创建了四个层次。使用定向梯度直方图（HOG）和局部二值模式（LBP）进行特征生成。通过基于ReliefF和邻域成分分析（NCA）的两级特征选择器（RFNCA）选择生成并连接的特征中最具判别力的特征。为了强调基于示例金字塔模型的特征生成的成功，选择了四种传统分类器进行分类，并且所提出的示例金字塔模型通过留一法交叉验证（LOOCV）实现了99.12%的分类准确率。结果和测试清楚地说明了基于示例金字塔模型的肱骨骨折分类方法的成功。结果还表明，所提出的示例金字塔模型实现的分类率高于肩部专科的骨科医生。
