Automatic detection of humerus bone fracture from x ray images using transfer learning approach

Abstract

Identification of bone fractures with computer aided detection and newlinediagnosis is an utmost need of today. It helps the radiologists in saving the time newlineand improving the performance. There were many image processing techniques newlineused earlier for detecting the bone fractures. Deep learning model in specific newlineconvolutional neural networks are widely used currently in the medical image newlineprocessing. It also extends its horizon in bone fracture detection from the X-Ray newlineimages. The commonly used dataset for bone fracture detection is MURA newlineDataset. This work concentrates on classifying the bone fractures and it employs newlinethe MURA dataset. While almost all the works that detects bone fracture newlineemploys the entire MURA Dataset, this work uses only the specific type of bone newlinecalled as the humerus bones and tries to classify the bone fractures. The work is newlinethree-fold, the first phase of the work concentrates on comparing the pre-trained newlinemodels DenseNet169 Model and VGG Model. Two variants of the DenseNet169 newlineModels one with pre-trained weights and another without the pre-trained weights newlineis experimented and the results are compared. The second phase of the work concentrates on comparing the three models VGG, DenseNet169 and DenseNet121. Two datasets, humerus dataset as it is in the MURA Dataset and the same one without the images that has metal newlinefitted to it. The performances of the models are tested and DenseNet169 is newlinefurther chosen for transfer learning process. The method of transfer learning newlineapproach used is carefully designed to meet the requirements. Customization of newlinethe layers was made with the consideration that the pre -trained models are newlinetrained with the ImageNet dataset and the dataset to which it has to be applied is newlineX-Ray images. The model is tested with both the dataset and the results are newlineinferred. The third phase of the work concentrates on developing a new CNN newlineModel. A customized model is developed for detecting the bone fracture of newlinehumerus bones. The model is tested with the two datasets. A slight change has newlinebeen made in the validation set for the second dataset where the X-Ray images newlinewith metals fixed to it are removed which is not the case with the first dataset. It newlinehas been observed that the newly designed Convolutional Neural Network model newlineperforms well than the model to which the transfer learning is applied as well as newlinethe pre-trained models DenseNet121, DenseNet169 and VGG16 models newline newline