(Tuesday, 7/16/2019) 10:30 AM - 11:00 AM
Room: Exhibit Hall | Forum 8
Purpose: We established a feasible digital X-ray photography classification model, based on the machine learning approach of Convolutional Neural Networks. We made a distinction between different radiotherapy patients, which can distinguish single patient’s X-Ray photography according to the image information.
Methods: We have selected 40-70 chest radiographs per person of ten patients at different times from ChestX-ray8 database (https://nihcc.app.box.com/v/ChestXray-NIHCC)provided by National Institutes of Health (NIH). We divided it into ten parts; nine parts as the training set and one is test set. In training set, 20% data was as the validation data, while 80% was the training data. After image size normalization processing, we import data into resnet-50 machine learning model. By adjusting Iteration times, we have enhanced the training effect and reduced the loss value of neural network. To assess the classification performance and test for overfitting, we performed 10-fold cross-validation.
Results: The trained model worked well on classification task: the accuracy can be up to 92%. Machine can find all X-ray images of the same person accurately. Furthermore, the performance index of the model under ten-fold cross validation is 81.8695%.
Conclusion: Machine learning can distinguish chest radiographs of different patients very well, such as identifying images of specific patient in large databases, even if the images are from different times. This kind of image recognition technology for specific patient images has potential research value in the fields of radiotherapy, radiography and the others. The next step of our task is expanding the class of classification, extending the scope of work to CT, mr and other modalities and develop patient recognition in other modality medical images. Meanwhile, in the image of radiotherapy, some work related to patient identification will be carry out.
Keywords
Cross Validation, Chest Radiography, Classifier Design
Taxonomy
Not Applicable / None Entered.
Contact Email
