Click here to


Are you sure ?

Yes, do it No, cancel

Radiography View Identification with Deep Learning

D Huo1*, L Harris2 , (1) University Colorado Denver, School of Medicine, Aurora, CO, (2) University of Colorado Hospital, Aurora, CO


(Wednesday, 7/17/2019) 10:30 AM - 11:00 AM

Room: Exhibit Hall | Forum 2

Purpose: Radiography has been clinically used for more than 100 years, and x-ray “views� have been standardized in technologist education, radiology order and billing code. The information of each “view�, including anatomy, patient position, and image orientation, is usually recorded in the EMR (Electronic Medical Record) and DICOM header. However, mismatch between the record and the actual image still happens quite often for various reasons, which represents possible wrong exams performed. In this abstract, deep learning models were trained and applied to a large in-house radiography image dataset to automatically identify the radiography image views from the image contents.

Methods: With the permission of local IRB, a large x-ray image dataset was established (total of 15141 images) for the most “popular� 143 radiography views used in our facility. Each image was labeled by an experienced radiographic technologist, to four levels: Level 1 (Abdomen, Chest, Knee, etc), Level 2 (Left, Right, etc), Level 3 (AP, Lateral, Oblique, etc) and Level 4 (Upright, Standing, etc). Based on these 4 levels of labels, the x-ray images are separated into 22, 38, 101, and 143 categories for level 1, 2, 3, and 4 respectively. A pre-trained Inception V3 network was modified to perform the classification task. For each category, 80% of the images were used for training, and 20% were used for validation. Separate models were trained and validated for each level of classification task.

Results: The pre-trained deep learning network with transferred learning could effectively classify the x-ray “views’ in each level. The accuracy of classification is 96%, 94.4%, 90.7% and 86.7% for level 1, 2, 3 and 4, respectively.

Conclusion: Deep learning could effectively classify the x-ray “views� at different levels and show great potential to help with the quality control in radiology.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email