Prediction of radiographic abnormalities by the use of bag-of-features and convolutional neural networks

This study evaluated the feasibility of bag-of-features (BOF) and convolutional neural networks (CNN) for computer-aided detection in distinguishing normal from abnormal radiographic findings. Computed thoracic radiographs of dogs were collected. For the purposes of this study, radiographic findings were used to distinguish between normal and abnormal in the following areas: (1) normal cardiac silhouette vs. cardiomegaly, (2) normal lung vs. abnormal lung patterns, (3) normal mediastinal position vs. mediastinal shift, (4) normal pleural space vs. pleural effusion, and (5) normal pleural space vs. pneumothorax. Images for training and testing the models consisted of ventrodorsal and lateral projection images of the same scale. The number of images used for each finding are as follow: 3142 for cardiomegaly (1571 normal and 1571 abnormal from 1143 dogs), 2086 for lung pattern (1043 normal and 1043 abnormal from 1247 dogs), 892 for mediastinal shift (446 normal and 446 abnormal from 387 dogs), 940 for pleural effusion (470 normal and 470 abnormal from 284 dogs), and 78 for pneumothorax (39 normal and 39 abnormal from 61 dogs). All data samples were divided so that 60% would be used for training the algorithms and 40% for testing the two models. The performance of the classifiers was evaluated by calculating the accuracy, sensitivity and specificity. The accuracy of both models ranged from 79.6% to 96.9% in the testing set. CNN showed higher accuracy (CNN; 92.9-96.9% and BOF; 79.6-96.9%) and sensitivity (CNN; 92.1-100% and BOF; 74.1-94.8%) than BOF. In conclusion, both BOF and CNN have potential to be useful for improving work efficiency by double reading.