Previous Abstract | Next Abstract
Printable Version
A1060
October 13, 2018
10/13/2018 10:30:00 AM - 10/13/2018 12:30:00 PM
Room North, Hall D, Area C
Development of Medical Ampule Label Detection and Identification Device with Neural Network Deep Learning Technology
Yusuke Kasuya, M.D.,Ph.D., Hidetsugu Asano, Ph.D., Shota Moriwaki, M.D., Keiko Okuyama, M.D., M.S., Yoshihiro Muragaki, M.D.,Ph.D., Makoto Ozaki, M.D.,Ph.D.
Tokyo Women's Medical University, Tokyo, Japan
Disclosures: Y. Kasuya: None. H. Asano: None. S. Moriwaki: None. K. Okuyama: None. Y. Muragaki: None. M. Ozaki: None.
Introduction: Drug administration errors frequently happen in the operating room, which may result in a serious accident. Because it is not feasible to perform routine two-person double checks in daily practice in the operating room, we need other strategies to prevent drug errors, and a computer-device-assisted double check can be one of the solutions. The aim of this project is to develop a novel medical ampule identification system using a deep learning convolutional neural network model. Methods: This study did not require an institutional research board approval as no humans or animals were involved. Seventy-one different ampules were video recorded, and then 30 still images were captured from the video while rotating the ampule 360 degree around its long axis at an interval of 12 degree. The object identification model was developed using YOLO9000 (https://pjreddie.com/darknet/yolo/), which is a novel promising object detection and classification application framework based on a deep neural network model. We built a 22-layer convolutional neural network referring DarkNetTM library, and trained it with a total of 64800 images. The image data were normalized into 416 by 416 pixels and 255-step RGB color gradation. Deep learning hyper-parameters were adjusted as the activation function of Leaky ReLU, the batch size of 64, the optimizer of ADAM, and the learning rate of 0.0001. Before each convolutional layer, batch normalization was applied. Deep learning was processed with Windows 10 Pro 64 bit, Intel Intel® Core™ i7-5930K, 32 GB memory and NVIDIA® TITAN X (Pascal™).Results and Discussions: We firstly developed to the system to identify 15 ampules to adjust hyper parameters, and then finally expanded into a 71 different ampule-identification system. Ampule labels were detected and surrounded by rectangle with the identified medicine (Figure 1). The combined identification accuracy matrix in the preliminary model was shown in Figure 2. All 71 ampules were successfully identified with prompt response at a frequency of 30 Hz. To evaluate the accuracy of the final model, new 129600 images which had not been used for developing the model[KO1] were tested. The precision and recall were 92.9% and 79.1% respectively. Our new system may be able to apply to a device for computer-assisted double checking. Although a barcode system has been widely used to improve drug administration safety, it cannot detect a label without barcode. Our system has potential to identify labels peeled off from the ampule and stuck on a syringe without barcode. It can also work with any video camera without scanning the label in close distance. Although all ampules were successfully identified with high accuracy, further examination will be required to prevent overfitting, which may reduce the reliability of the system. Non-specific information on the label (e.g. expiry date, production identification number) can cause overfitting. We plan to expand this system to apply for all medicines in the operating room.

Figure 1
Figure 2

Copyright © 2018 American Society of Anesthesiologists